benchmark_id,benchmark_name,category,metric,num_problems,source_url,canonical_setting_json
aa_intelligence_index,AA Intelligence Index,Composite,index score,12826.0,https://artificialanalysis.ai/methodology/intelligence-benchmarking,"{""higher_is_better"":true,""judge"":""mixed scoring protocols"",""metric_type"":""index"",""multimodal_input"":false,""notes"":""Composite weighted index over 10 evaluations. Count is actual model generations across official questions/tasks and repeats: GDPval-AA 220*1, tau2-Bench Telecom 114*3, Terminal-Bench Hard 44*3, SciCode 288*3, AA-LCR 100*3, AA-Omniscience 6000*1, IFBench 294*5, HLE text-only 2158*1, GPQA Diamond 198*5, CritPt 70*5 = 12826. Cost burden is heterogeneous; tools=composite intentionally avoids applying one agentic multiplier to every component."",""range"":null,""sampling"":""included in num_problems"",""tools"":""composite"",""version"":""Artificial Analysis Intelligence Index v4.0.4 (March 2026)""}"
aa_lcr,AA Long Context Reasoning,Long Context,% correct,300.0,https://artificialanalysis.ai/methodology/intelligence-benchmarking,"{""harness"":""official Artificial Analysis LCR"",""higher_is_better"":true,""judge"":""official AA equality checker"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official AA-LCR has 100 open-answer questions over roughly 100k-token document contexts and runs three repeats, so num_problems records 300 physical model generations. StepFun's avg@16 observation is a score-level repeated-sampling setting."",""range"":[0,100],""sampling"":""3 repeats per question; pass@1 aggregated"",""tools"":""none"",""version"":""Artificial Analysis Long Context Reasoning, 100 questions""}"
aethercode,AetherCode,Coding,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""AetherCode""}"
agentcompany,AgentCompany,Agentic,%,,https://huggingface.co/MiniMaxAI/MiniMax-M2,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per MiniMax M2 model card."",""range"":[0,100],""tools"":""agentic"",""version"":""AgentCompany""}"
ai2d,AI2D,Multimodal,%,,https://mistral.ai/news/mistral-medium-3,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Mistral Medium 3 blog: AI2 Diagram understanding benchmark, 0-shot."",""range"":[0,100],""tools"":""none"",""version"":""AI2D""}"
aider_polyglot_diff,Aider Polyglot (diff mode),Coding,%,450.0,https://aider.chat/2024/12/21/polyglot.html,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Aider Polyglot uses 225 selected Exercism coding tasks across C++, Go, Java, JavaScript, Python, and Rust. The displayed leaderboard score corresponds to the second-try/pass_rate_2 setting, so cost count records actual model generations: 225 tasks times two tries = 450. Diff mode is selected by edit_format=diff."",""range"":[0,100],""sampling"":""included in num_problems"",""tools"":""agentic"",""version"":""Aider Polyglot benchmark; diff edit format""}"
aider_polyglot_whole,Aider Polyglot (whole mode),Coding,%,450.0,https://aider.chat/2024/12/21/polyglot.html,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Aider Polyglot uses 225 selected Exercism coding tasks across C++, Go, Java, JavaScript, Python, and Rust. The displayed leaderboard score corresponds to the second-try/pass_rate_2 setting, so cost count records actual model generations: 225 tasks times two tries = 450. Whole mode is selected by edit_format=whole."",""range"":[0,100],""sampling"":""included in num_problems"",""tools"":""agentic"",""version"":""Aider Polyglot benchmark; whole edit format""}"
aime_2024,AIME 2024,Math,% correct (pass@1),30.0,https://artofproblemsolving.com/wiki/index.php/2024_AIME,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""tools"":""none"",""version"":""AIME-2024-I+II (30 problems)""}"
aime_2025,AIME 2025,Math,% correct (pass@1),30.0,https://artofproblemsolving.com/wiki/index.php/2025_AIME,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""tools"":""none"",""version"":""AIME-2025-I+II (30 problems)""}"
aime_2026,AIME 2026,Math,% correct (pass@1),30.0,https://huggingface.co/datasets/MathArena/aime_2026,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false. Canonical row combines AIME 2026 I and II: 30 problems total."",""range"":[0,100],""tools"":""none"",""version"":""AIME-2026-I+II (30 problems)""}"
ainstein_bench,AInsteinBench,Science Discovery,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""AInsteinBench""}"
all_angles,All-Angles,Vision Spatial,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""All-Angles""}"
alpacaeval_2,AlpacaEval 2.0 (LC-winrate),Chat,%,,https://arxiv.org/abs/2501.12948,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per DS R1 paper."",""range"":[0,100],""tools"":""none"",""version"":""AlpacaEval 2.0 (LC-winrate)""}"
apex_agents,APEX-Agents,Agentic,,,https://deepmind.google/models/evals-methodology/gemini-3-pro,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""DeepMind APEX-Agents long-horizon professional benchmark. Distinct from MathArena Apex 2025."",""range"":[0,100],""version"":""APEX-Agents (long-horizon professional tasks)""}"
apex_shortlist,Apex Shortlist,Math,% correct (pass@1),,https://matharena.ai/apex/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""tools"":""none"",""version"":""Apex shortlist""}"
arc_agi_1,ARC-AGI-1,Reasoning,% correct,400.0,https://arcprize.org/arc-agi/1/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""tools"":""none"",""version"":""ARC-AGI-1 (semi-private 400)""}"
arc_agi_2,ARC-AGI-2,Reasoning,% correct,120.0,https://arcprize.org/arc-agi/2/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""ARC-AGI-2 v2 semi-private evaluation tier contains 120 calibrated tasks. Each task passes only when all test grids are exact; up to two outputs per test input are allowed. Tool/scaffold differences remain cell settings."",""range"":[0,100],""tools"":""none"",""version"":""ARC-AGI-2 v2 semi-private evaluation set (120 tasks)""}"
arc_challenge,ARC Challenge,Reasoning,% accuracy,1172.0,https://huggingface.co/datasets/allenai/ai2_arc/resolve/210d026faf9955653af8916fad021475a3f00453/README.md,"{""higher_is_better"":true,""judge"":""answer-key accuracy"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official immutable AI2 dataset card reports 1,172 test questions. Few-shot count is observation-specific."",""range"":[0,100],""sampling"":""one multiple-choice response per question"",""tools"":""none"",""version"":""AI2 ARC-Challenge test split""}"
arcagi1_image,ArcAGI1-Image,Vision Puzzles,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""ArcAGI1-Image""}"
arcagi2_image,ArcAGI2-Image,Vision Puzzles,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""ArcAGI2-Image""}"
arena_hard,Arena-Hard Auto,Instruction Following,% win rate,500.0,https://lmarena.ai/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""tools"":""none"",""version"":""Arena-Hard-Auto""}"
artifactsbench,ArtifactsBench,Coding,%,5475.0,https://github.com/Tencent-Hunyuan/ArtifactsBenchmark,"{""higher_is_better"":true,""judge"":""Gemini-2.5-Pro MLLM-as-Judge with checklist-guided scoring"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Official ArtifactsBench contains 1825 diverse tasks / HF rows. The MiniMax-M2 score source reports scores averaged over three runs with the official implementation and stable Gemini-2.5-Pro judge. Count records actual model generations for the BenchPress row: 1825 tasks times three runs = 5475. Evaluation renders generated artifacts, captures dynamic behavior, and scores visual/interactivity quality with a multimodal judge."",""range"":[0,100],""sampling"":""included in num_problems"",""tools"":""none for model; evaluator renders generated artifacts and captures screenshots"",""version"":""ArtifactsBench full benchmark; MiniMax-M2 reported setting""}"
babe,BABE,Reasoning,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""BABE""}"
babyvision,BabyVision,Multimodal,% accuracy,388.0,https://huggingface.co/datasets/UnipatAI/BabyVision,"{""higher_is_better"":true,""judge"":""LLM judge compares model output to ground truth answer"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Official BabyVision MLLM evaluation has 388 visual reasoning tasks; BabyVision-Gen is a separate generation-track benchmark."",""range"":[0,100],""sampling"":""pass@1"",""version"":""BabyVision MLLM evaluation""}"
beyond_aime,Beyond AIME,Math,%,100.0,https://huggingface.co/datasets/ByteDance-Seed/BeyondAIME,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""HF dataset card reports one test split with 100 problems; answers are positive integers with automated exact verification. Per Seed-Thinking-v1.5 paper."",""range"":[0,100],""tools"":""none"",""version"":""Beyond AIME""}"
bfcl,BFCL,Tool use,,,https://cohere.com/research/papers/command-a-technical-report.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Function calling benchmark. Distinct from bfcl_v3."",""range"":[0,100],""version"":""Berkeley Function Calling Leaderboard (Tau-bench predecessor)""}"
bfcl_v3,BFCL v3,Tool use,,,https://gorilla.cs.berkeley.edu/leaderboard.html,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Function-calling benchmark, FC format"",""range"":[0,100],""version"":""BFCL v3 (Berkeley Function Calling Leaderboard)""}"
bfcl_v3_multiturn,BFCL v3 (Multi-Turn),Tool Use,%,,https://huggingface.co/deepseek-ai/DeepSeek-R1-0528,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per DeepSeek R1-0528 model card."",""range"":[0,100],""tools"":""agentic"",""version"":""BFCL v3 (Multi-Turn)""}"
bfcl_v4,BFCL v4,Tool Use,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""tool calls"",""version"":""BFCL v4""}"
bigbench_extra_hard,BigBench Extra Hard,Reasoning,micro accuracy (%),4520.0,https://github.com/google-deepmind/bbeh,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Full 4,520-example benchmark; Gemma reports example-weighted micro-average accuracy."",""range"":[0,100],""tools"":""none"",""version"":""Big-Bench Extra Hard full benchmark""}"
bigbench_hard,BigBench Hard (BBH),Reasoning,% exact-match accuracy,6511.0,https://github.com/suzgunmirac/BIG-Bench-Hard/tree/9ee07bd481feebf959a6b59d61ea57bdcf30964d,"{""higher_is_better"":true,""judge"":""task-specific exact-match normalization"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The official paper calls BBH 23 challenging tasks; the locked release contains 27 JSON task files and exactly 6,511 prompt examples. Count is actual model generations."",""range"":[0,100],""sampling"":""one response per released prompt"",""tools"":""none"",""version"":""BIG-Bench Hard immutable official release""}"
bigcodebench,BigCodeBench,Coding,pass@1 %,1140.0,https://bigcode-bench.github.io/,"{""higher_is_better"":true,""judge"":""sandboxed unit-test evaluator"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official BigCodeBench complete/instruct evaluation uses generated code executed by the benchmark sandbox and unit tests; no agentic tools. Score observations must identify complete versus instruct split."",""range"":[0,100],""tools"":""none"",""version"":""BigCodeBench (1140 full set)""}"
biobench,BIObench,Science Discovery,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""BIObench""}"
bird_sql,Bird-SQL (Dev),Coding,,,https://bird-bench.github.io/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Natural language to executable SQL on Bird-SQL dev split."",""range"":[0,100],""version"":""Bird-SQL Dev split (NL\u2192SQL)""}"
bixbench,BixBench Zero-Shot MCQ,Science,accuracy (%),205.0,https://github.com/Future-House/BixBench,"{""harness"":""BixBench official zero-shot MCQ; score-level agent harness"",""higher_is_better"":true,""judge"":""zero-shot multiple-choice accuracy"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Public benchmark contains 205 computational-biology questions. The xAI score uses Grok Build with analysis tools enabled by default."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""analysis tools available"",""version"":""BixBench zero-shot MCQ""}"
blink,BLINK,Vision Spatial,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""BLINK""}"
browsecomp,BrowseComp,Agentic,accuracy (%),1266.0,https://raw.githubusercontent.com/openai/simple-evals/652c89d0ca9df547706735883097e9537d40dc47/browsecomp_eval.py,"{""harness"":""source-reported browser scaffold"",""higher_is_better"":true,""judge"":""official BrowseComp grading protocol"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The locked official dataset contains 1,266 questions. Browser scaffold and context management remain score-level settings."",""range"":[0,100],""sampling"":""one answer per question"",""tools"":""web browsing agent"",""version"":""BrowseComp official 1,266-question release""}"
browsecomp_cm,BrowseComp (w/ Context Manage),Agentic,accuracy (%),,https://z.ai/blog/glm-4.7,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Context management: discard-all strategy (not retain-5-turns). Per z.ai/blog/glm-4.7 and GLM-5.1 blog footnote."",""range"":[0,100],""tools"":""agentic"",""version"":""BrowseComp with discard-all context management""}"
browsecomp_long_context_128k,BrowseComp Long Context 128k,Long Context,% accuracy,1266.0,https://openai.com/index/gpt-5-1-for-developers/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""OpenAI GPT-5.1 appendix reports BrowseComp Long Context 128k but does not publish a separate count. Use the official BrowseComp 1,266-row test set as the source-backed count unless a 128k-specific slice is found."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none specified for the 128k long-context row"",""version"":""BrowseComp Long Context 128k""}"
browsecomp_long_context_256k,BrowseComp Long Context 256k,Long Context,,,,"{""judge"":""rule-based"",""notes"":""Per OpenAI GPT-5 developer blog https://openai.com/index/introducing-gpt-5-for-developers/""}"
browsecomp_zh,BrowseComp-ZH,Agentic search,,1156.0,https://github.com/PALIN2018/BrowseComp-ZH,"{""higher_is_better"":true,""judge"":""LLM-assisted answer extraction / grading"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""BrowseComp-ZH official paper and repository define 289 native-Chinese multi-hop web-browsing questions across 11 domains. The Moonshot/Kimi score source reports BrowseComp-ZH with avg@4, so the cost count records actual model generations: 289 questions times 4 independent runs = 1,156. Do not use the parent English BrowseComp count."",""range"":[0,100],""sampling"":""included in num_problems"",""tools"":""web browsing and search tools"",""version"":""BrowseComp-ZH official 289-question benchmark; Moonshot avg@4 setting""}"
brumo_2025,BRUMO 2025,Math,% correct (pass@1),30.0,https://huggingface.co/datasets/MathArena/brumo_2025,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""sampling"":""samples=4"",""tools"":""none"",""version"":""BRUMO 2025""}"
bullshit_pushback,Bullshit-Bench (Clear Pushback),Behavior,% clear pushback,55.0,https://github.com/petergpt/bullshit-benchmark,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""tools"":""none"",""version"":""Bullshit-pushback (55)""}"
c_eval,C-Eval (Chinese),Knowledge,%,12342.0,https://huggingface.co/datasets/ceval/ceval-exam,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""HF dataset card reports 13,948 total questions across splits; the test split has 12,342 scored multiple-choice questions across 52 subjects."",""range"":[0,100],""tools"":""none"",""version"":""C-Eval (Chinese)""}"
cgbench,CGBench,Video,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""CGBench""}"
chartqa,ChartQA,Multimodal,%,,https://mistral.ai/news/mistral-medium-3,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Mistral Medium 3 blog: Chart visual question answering, 0-shot."",""range"":[0,100],""tools"":""none"",""version"":""ChartQA""}"
chartqapro,ChartQAPro,Multimodal,overall answer accuracy (%),1948.0,https://arxiv.org/abs/2504.05506,"{""harness"":""official"",""higher_is_better"":true,""judge"":""answer-type-aware official parser/evaluator"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""1,948 questions over 1,341 charts."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""ChartQAPro""}"
charxiv_descriptive,CharXiv Descriptive,Vision,% accuracy,4000.0,https://charxiv.github.io/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Official leaderboard validation set has 1,000 charts and 5,000 questions; HF schema has four descriptive question fields per chart, so descriptive evaluation is 4,000 model answers."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""CharXiv validation descriptive questions""}"
charxiv_reasoning,CharXiv Reasoning,Multimodal,% accuracy,1000.0,https://charxiv.github.io/,"{""higher_is_better"":true,""judge"":""gpt-4o-2024-05-13, temperature=0, seed=42, binary answer-key judge"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""CharXiv v1.0 validation reasoning subset has 1,000 charts and one reasoning answer per chart. Official evaluator uses gpt-4o-2024-05-13 at temperature 0 and seed 42."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""CharXiv validation reasoning questions""}"
chatbot_arena_elo,Chatbot Arena Elo,Human Preference,Elo rating,8000.0,https://arxiv.org/abs/2403.04132,"{""higher_is_better"":true,""judge"":""human pairwise preference votes"",""metric_type"":""elo"",""multimodal_input"":false,""notes"":""Live crowdsourced pairwise comparison benchmark. The paper reports over 240K votes total and about 8K votes per model on average as of Jan 2024; use 8K battles as the source-backed per-model cost proxy. No fixed static item set."",""range"":null,""tools"":""none"",""version"":""LMArena Chatbot Arena live Elo, text-only general leaderboard""}"
chinese_simpleqa,Chinese-SimpleQA,Knowledge,%,3000.0,https://huggingface.co/datasets/OpenStellarTeam/Chinese-SimpleQA,"{""higher_is_better"":true,""judge"":""LLM grader"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Protocol audit: short Chinese factual QA. Each item asks a short-answer factual question; model output is judged for correctness against reference answers. HF dataset card reports 3,000 questions across 6 topics and says grading is run via existing LLMs. No tools, multimodal input, long context, multi-turn interaction, or repeated sampling is specified."",""range"":[0,100],""sampling"":""single-pass; no repeated sampling specified"",""tools"":""none"",""version"":""Chinese-SimpleQA""}"
cl_bench,CL-Bench,Long Context,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""CL-Bench""}"
claw_eval_pass3,Claw Eval (pass^3),Agentic,all-three-pass rate (%),597.0,https://raw.githubusercontent.com/claw-eval/claw-eval/5680b8b11ff2ee5dd2b07b89086a29a5c5c984d7/README.md,"{""harness"":""official Claw-Eval v1.1"",""higher_is_better"":true,""judge"":""full-trajectory completion/safety/robustness grading"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""This campaign identity is the non-multimodal aggregate: 161 general plus 38 multi-turn tasks. Pass^3 requires all three trials, so num_problems records exactly 597 physical trajectories. The separate 101-task multimodal section is not included in this benchmark identity."",""range"":[0,100],""sampling"":""3 independent successful trajectories per task"",""tools"":""official Claw-Eval agent environment"",""version"":""Claw-Eval v1.1 non-multimodal Pass^3""}"
cluewsc,CLUEWSC,Chinese,%,2574.0,https://huggingface.co/datasets/clue/clue,"{""higher_is_better"":true,""judge"":""rule-based"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Protocol audit: Chinese Winograd/coreference-style binary classification. Each item contains a Chinese text and two target spans; the model predicts true/false and scoring is exact match/accuracy against the class label. HF clue/clue dataset card reports cluewsc2020 splits with 2,574 test examples, 1,244 train examples, and 304 validation examples. No LLM judge, tools, multimodal input, long context, multi-turn interaction, or repeated sampling is used."",""range"":[0,100],""sampling"":""single-pass"",""tools"":""none"",""version"":""CLUEWSC""}"
cmimc_2025,CMIMC 2025,Math,% correct (pass@1),40.0,https://huggingface.co/datasets/MathArena/cmimc_2025,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""sampling"":""samples=4"",""tools"":""none"",""version"":""CMIMC 2025""}"
cmmlu,CMMLU (Chinese),Knowledge,% accuracy,11582.0,https://huggingface.co/datasets/haonan-li/cmmlu/resolve/efcc940752ea4a1ea94d2727f11f83858d64fc8e/README.md,"{""higher_is_better"":true,""judge"":""answer-key accuracy"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The locked official v1.0.1 archive contains 11,582 test questions across 67 subjects; the 5-question dev split is used for the source's reported 5-shot prompting."",""range"":[0,100],""sampling"":""one multiple-choice response per question"",""tools"":""none"",""version"":""CMMLU v1.0.1 test split, 67 subjects""}"
cnmo_2024,CNMO 2024,Math,%,6.0,https://www.cms.org.cn/Home/comp/comp_details/id/1253.html,"{""higher_is_better"":true,""judge"":""rule-based"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Protocol audit: Chinese National High School Mathematics Olympiad 2024 finals, pure text olympiad math. The official CMS page identifies the 2024 national final / 40th winter camp; the standard CMO format is two days with 3 problems per day (format source: https://zh.wikipedia.org/wiki/\u4e2d\u56fd\u6570\u5b66\u5965\u6797\u5339\u514b), so the scored set has 6 proof-style math problems. DeepSeek-R1-0528 reports CNMO 2024 as Pass@1 and states that benchmarks requiring sampling use temperature 0.6, top-p 0.95, and 16 responses per query to estimate pass@1. Scoring is rule-based/manual exact mathematical correctness; no LLM judge, tools, multimodal input, long context, or multi-turn interaction."",""range"":[0,100],""sampling"":""samples=16"",""tools"":""none"",""version"":""CNMO 2024""}"
codeforces_avg8,Codeforces (avg@8),Coding,%,,https://arxiv.org/abs/2504.13914,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Seed-Thinking-v1.5 paper."",""range"":[0,100],""tools"":""none"",""version"":""Codeforces (avg@8)""}"
codeforces_pass8,Codeforces (pass@8),Coding,%,,https://arxiv.org/abs/2504.13914,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Seed-Thinking-v1.5 paper."",""range"":[0,100],""tools"":""none"",""version"":""Codeforces (pass@8)""}"
codeforces_rating,Codeforces Rating,Coding,Elo rating,,https://codeforces.com/,"{""higher_is_better"":true,""metric_type"":""rating"",""multimodal_input"":false,""notes"":""tools=agentic (scaffold-defined). Preferred = official harness tools (bash/editor for SWE-bench, browser for OSWorld/BrowseComp, official APIs for tau-bench/MCPAtlas/Toolathlon, terminal for Terminal-Bench). Non-official scaffolds \u2192 matches_canonical=false."",""range"":null,""tools"":""agentic"",""version"":""Codeforces live rating""}"
codesimpleqa,CodeSimpleQA,Coding,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""CodeSimpleQA""}"
collie,COLLIE,Instruction Following,%,2080.0,https://arxiv.org/abs/2307.08689,"{""higher_is_better"":true,""judge"":""rule-based"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Protocol audit: constrained text generation benchmark. Each item renders a natural-language instruction from a formal COLLIE constraint structure; the model outputs free-form text, and scoring checks whether the generated text satisfies the target constraint. The COLLIE paper reports COLLIE-v1 has 2,080 instances across 13 constraint structures. The official repo documents evaluation via the constraint checker, so scoring is rule-based/programmatic rather than LLM-judged. BenchPress cells from OpenAI/Doubao reports use pass@1/single-response settings. No tools, multimodal input, long context, or multi-turn interaction is used."",""range"":[0,100],""sampling"":""pass@1; single response"",""tools"":""none"",""version"":""COLLIE""}"
complexfuncbench,ComplexFuncBench,Tool Use,%,1000.0,https://github.com/THUDM/ComplexFuncBench,"{""higher_is_better"":true,""judge"":""ComplexEval automatic matching plus final-response LLM evaluation"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official paper/repo define 1,000 samples: 600 single-domain and 400 cross-domain. Each sample is a multi-step function-calling dialogue; average 3.26 steps and 5.07 calls per sample. Includes real API responses and 128k long-context cases."",""range"":[0,100],""tools"":""function calling"",""version"":""ComplexFuncBench 128k long-context function calling""}"
contphy,ContPhy,Video,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""ContPhy""}"
corpusqa_1m,CorpusQA 1M,Long Context,%,,https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per DeepSeek V4-Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""CorpusQA 1M""}"
countbench,CountBench,Vision Counting,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""CountBench""}"
covost2,CoVoST2 (21 lang),Audio,,,https://github.com/facebookresearch/covost,"{""higher_is_better"":true,""metric_type"":""bleu"",""multimodal_input"":true,""notes"":""Automatic speech translation across 21 languages (BLEU score)."",""range"":[0,100],""version"":""CoVoST2 21-language speech translation (BLEU)""}"
creative_writing_v3,Creative Writing v3 (Elo Normalized),Creative,elo,,https://x.ai/news/grok-4-1,"{""higher_is_better"":true,""metric_type"":""elo"",""multimodal_input"":false,""notes"":""Creative Writing v3: 32 prompts \u00d7 3 iterations. LLM-judged with rubrics + pairwise battles. Elo normalized. Per xAI Grok 4.1 blog."",""range"":[1000,2000],""tools"":""none"",""version"":""Creative Writing v3 (Elo Normalized)""}"
critpt,CritPt,Science,% correct,70.0,https://huggingface.co/datasets/CritPt-Benchmark/CritPt,"{""higher_is_better"":true,""judge"":""automated rule-based scoring server"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Protocol audit: frontier research-level physics benchmark. The public test set has 70 challenges; the broader benchmark has 71 composite research challenges plus an example and 190 checkpoint tasks. Each challenge is a text-only, unpublished physics research problem spanning modern physics subfields, with guess-resistant, machine-verifiable answers. Primary leaderboard metric is average challenge accuracy over 5 runs x 70 test challenges. The official pipeline submits complete batches to an automated grading server customized for advanced physics-specific output formats. Canonical BenchPress setting is no tools; with-code/web-tool variants are non-canonical. The official repo's no-tool config disables Python and web search; reasoning-model examples use large reasoning budgets (e.g. 27k reasoning tokens)."",""range"":[0,100],""sampling"":""trials=5"",""tools"":""none"",""version"":""CRITPT""}"
crossvid,CrossVid,Video,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""CrossVid""}"
ctf_internal,Capture-the-Flags challenge tasks (Internal),Cyber,%,,https://openai.com/index/introducing-gpt-5-5/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Hardest CTF challenges from system cards plus additional hard challenges."",""range"":[0,100],""tools"":""agentic"",""version"":""Capture-the-Flags challenge tasks (Internal)""}"
cybench,Cybench,Cyber,%,40.0,https://arxiv.org/abs/2408.08926,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Public CTF benchmark: 40 challenges from 4 competitions (Zhang et al., 2024). Anthropic evaluated 39/40 (1 skipped due to infra/timing). Score = % of 39 attempted. Pass@30 trials."",""range"":[0,100],""tools"":""agentic"",""version"":""Cybench (public)""}"
cybergym,CyberGym,Agentic,% solved,1507.0,https://www.cybergym.io/,"{""higher_is_better"":true,""judge"":""PoC reproduced on vulnerable version and not on fixed version"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official benchmark has 1,507 historical vulnerability instances from 188 projects. Agents receive vulnerability description and unpatched codebase, generate PoCs, and are scored by execution against vulnerable/fixed program versions. The 10-task subset is not canonical."",""range"":[0,100],""tools"":""agentic code execution environment"",""version"":""CyberGym Level 1 vulnerability reproduction""}"
cybersecurity_ctf,Cybersecurity Capture The Flag Challenges,Cyber,%,,https://openai.com/index/introducing-gpt-5-3-codex/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Cybersecurity CTF benchmark per OpenAI GPT-5.3-Codex blog. Note: distinct from ctf_internal (GPT-5.5 blog uses different problem set)."",""range"":[0,100],""tools"":""agentic"",""version"":""Cybersecurity Capture The Flag Challenges""}"
da_2k,DA-2K,Vision Spatial,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""DA-2K""}"
deepconsult,DeepConsult,Deep Research,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""research tools"",""version"":""DeepConsult""}"
deepresearchbench,DeepResearchBench,Deep Research,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""research tools"",""version"":""DeepResearchBench""}"
deepsearchqa_acc,DeepSearchQA (Accuracy),Search Agent,accuracy (%),900.0,https://huggingface.co/datasets/google/deepsearchqa/tree/b2623f8653065c2672de6d941fc5434cd652376c,"{""harness"":""DeepSearchQA official evaluation"",""higher_is_better"":true,""judge"":""Gemini 2.5 Flash with the official Kaggle starter grading prompt"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The pinned official dataset has 900 prompts across 17 fields; changing autorater or prompt can significantly change results."",""range"":[0,100],""sampling"":""one answer per prompt"",""tools"":""web search agent"",""version"":""DeepSearchQA official 900-prompt accuracy (%)""}"
deepsearchqa_f1,DeepSearchQA (F1),Search Agent,F1 (%),900.0,https://huggingface.co/datasets/google/deepsearchqa/tree/b2623f8653065c2672de6d941fc5434cd652376c,"{""harness"":""DeepSearchQA official evaluation"",""higher_is_better"":true,""judge"":""Gemini 2.5 Flash with the official Kaggle starter grading prompt"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The pinned official dataset has 900 prompts across 17 fields; changing autorater or prompt can significantly change results."",""range"":[0,100],""sampling"":""one answer per prompt"",""tools"":""web search agent"",""version"":""DeepSearchQA official 900-prompt F1 (%)""}"
der2_bench,DeR2 Bench,Reasoning,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""DeR2 Bench""}"
disco_x,Disco-X,Multilingual,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""Disco-X""}"
docvqa,DocVQA,Multimodal,%,,https://mistral.ai/news/mistral-medium-3,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Mistral Medium 3 blog: Document visual question answering, 0-shot."",""range"":[0,100],""tools"":""none"",""version"":""DocVQA""}"
drop,DROP,Reasoning,%,9536.0,https://huggingface.co/datasets/EleutherAI/drop,"{""higher_is_better"":true,""judge"":""rule-based"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""DROP is passage-question reading comprehension requiring discrete reasoning. HF EleutherAI/drop reports 77,409 train rows and 9,536 validation rows; use validation as the scored evaluation split. HF ucinlp/drop reports 9,535 validation rows, so the one-row discrepancy is noted and EleutherAI/drop is used because it matches common lm-eval-style benchmark packaging. The official paper describes DROP as a 96k-question benchmark. Official evaluation uses normalized exact match and F1 over number/date/span answers; no LLM judge or tool use."",""range"":[0,100],""sampling"":""single-pass"",""tools"":""none"",""version"":""DROP validation split""}"
dude,DUDE,Document/Chart,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""DUDE""}"
dynamath,DynaMath,Math,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""DynaMath""}"
egoschema,EgoSchema (test),Video,,,https://egoschema.github.io/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Long-form egocentric video QA across multiple domains."",""range"":[0,100],""version"":""EgoSchema test split""}"
egotempo,EgoTempo,Video,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""EgoTempo""}"
emma,EMMA,Vision STEM,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""EMMA""}"
encyclo_k,Encyclo-K,Knowledge,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""Encyclo-K""}"
eq_bench3,"EQ-Bench3 (Emotional Intelligence, Elo Normalized)",EQ,elo,,https://x.ai/news/grok-4-1,"{""higher_is_better"":true,""metric_type"":""elo"",""multimodal_input"":false,""notes"":""EQ-Bench3: 45 roleplay scenarios \u00d7 3 turns. LLM-judged with rubrics + pairwise battles. Elo normalized. Per xAI Grok 4.1 blog."",""range"":[1000,2000],""tools"":""none"",""version"":""EQ-Bench3 (Emotional Intelligence, Elo Normalized)""}"
erqa,ERQA,Vision,%,400.0,https://github.com/embodiedreasoning/ERQA,"{""higher_is_better"":true,""judge"":""rule-based"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Official ERQA GitHub README says the full benchmark consists of 400 examples. Questions are multimodal interleaved images and text, phrased as multiple-choice questions, with answers provided as a single letter (A, B, C, D). The dataset covers embodied/spatial/trajectory/action/state-estimation reasoning for real-world robotics scenarios. HF mirrors GeorgeBredis/ERQA and FlagEval/ERQA both report 400 rows. Use exact letter accuracy; no LLM judge, tools, or agentic environment are part of the canonical benchmark."",""range"":[0,100],""sampling"":""single-pass"",""tools"":""none"",""version"":""ERQA full benchmark""}"
expert_swe,Expert-SWE (Internal),Coding,%,,https://openai.com/index/introducing-gpt-5-5/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Internal OpenAI software engineering benchmark."",""range"":[0,100],""tools"":""agentic"",""version"":""Expert-SWE (Internal)""}"
facts_benchmark,FACTS Benchmark Suite,Factuality,,,https://deepmind.google/models/gemini/flash/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Factuality across grounding, parametric, search, and multimodal."",""range"":[0,100],""version"":""FACTS Benchmark Suite (grounding/parametric/search/MM)""}"
facts_grounding,FACTS Grounding,Factuality,,1719.0,https://arxiv.org/abs/2501.03200,"{""higher_is_better"":true,""judge"":""LLM judge ensemble (Gemini 1.5 Pro, GPT-4o, Claude 3.5 Sonnet)"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""FACTS Grounding evaluates whether long-form model responses are factually accurate and grounded in a provided context document. The paper reports 1,719 total examples split into Open N=860 and Blind N=859. Each prompt includes a user request and a full document, with context up to 32k tokens. Models generate long-form responses; scoring uses prompted LLM judges in two phases: responses are first disqualified if they do not fulfill the user request, then judged accurate if fully grounded in the document. The factuality score aggregates three judge models (Gemini 1.5 Pro, GPT-4o, Claude 3.5 Sonnet) to mitigate judge bias."",""range"":[0,100],""sampling"":""single-pass model response; scored by three prompted judge models plus eligibility filter"",""tools"":""none"",""version"":""FACTS Grounding long-context factuality benchmark""}"
factscore,FActScore (hallucination rate),Hallucination,%,500.0,https://github.com/shmsw25/FActScore,"{""higher_is_better"":false,""judge"":""retrieval+LLM judge/factuality estimator"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official FActScore evaluates long-form biography generation for factual precision. The README defines two prompt-entity sets: 183 labeled entities for human-annotated sections and 500 unlabeled entities for broad model evaluation; use the 500-entity unlabeled set as the scored benchmark count. Each model generates a biography for a person entity, then FActScore decomposes the generation into atomic facts and verifies each fact against a Wikipedia knowledge source using retrieval+ChatGPT or retrieval+LLAMA+NP. The official README estimates API cost at about $1 per 100 sentences and reports that 6,500 generations from 13 LMs would have cost $26K if evaluated by humans. Some provider tables report hallucination rate (lower is better) rather than FActScore factual precision (higher is better); preserve source-level score semantics in score-cell notes."",""range"":[0,100],""sampling"":""single-pass generation; each biography is decomposed into atomic facts and verified against Wikipedia"",""tools"":""none"",""version"":""FActScore unlabeled 500-entity biography set; benchmark tables may use either FActScore or hallucination rate""}"
finance_agent,Finance Agent v1.1,Agentic,% solved,537.0,https://arxiv.org/abs/2508.00828,"{""higher_is_better"":true,""judge"":""LLM-as-judge rubric and contradiction grader"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Finance Agent Benchmark evaluates autonomous finance agents on expert-authored real-world financial analysis questions requiring recent SEC filings and open-web information. The paper reports 537 expert-authored questions across nine task categories; each entry includes a question, ground-truth answer, source documents, and step-by-step solution approach, and all reported metrics were calculated on the complete 537 samples. The public/private/test split is 50/150/337, but the paper's benchmark results use all 537 samples. The harness gives models Google Search, EDGAR search, HTML parsing, and retrieved-document tools. Scoring uses an LLM-as-judge rubric system: GPT-4o-generated rubrics are manually reviewed, contradiction rubrics check conflicts with the expert answer, and reported metrics include class-balanced accuracy and naive accuracy; figures default to class-balanced accuracy unless otherwise specified. Anthropic Opus 4.7 blog reports Finance Agent v1.1 scores, while the arXiv benchmark paper supplies the source-backed task count and protocol."",""range"":[0,100],""sampling"":""single evaluated agent run per question; paper reports class-balanced accuracy and naive accuracy"",""tools"":""agentic financial-analysis harness with GoogleSearch, EdgarSearch, ParseHTML, and RetrieveInformation tools"",""version"":""Finance Agent Benchmark v1.1; full 537-sample evaluation""}"
finsearchcomp,FinSearchComp,Search Agent,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""search"",""version"":""FinSearchComp""}"
finsearchcomp_global,FinSearchComp-Global,Search Agent,%,317.0,https://arxiv.org/abs/2509.13160,"{""higher_is_better"":true,""judge"":""LLM-as-a-Judge with task-specific rubrics"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""FinSearchComp is an open-domain financial search and reasoning benchmark. The paper reports 635 total expert-curated questions across Global and Greater China subsets; Figure 4 gives the Global subset task counts as T1=110, T2=119, and T3=88, so FinSearchComp-Global has 317 scored questions. Each question requires external search/tool use and has a single objective answer. Scoring uses LLM-as-a-Judge with task-specific rubrics and a binary 0-1 judgment; the paper reports roughly 95% agreement with human-verified labels on a representative validation sample. MiniMax-M2 reports FinSearchComp-global scores from the open-source FinSearchComp framework using both search and Python tools."",""range"":[0,100],""sampling"":""single evaluated answer per question; 0-1 correctness"",""tools"":""open-domain search agent; MiniMax-M2 reports use the open-source FinSearchComp framework with search and Python tools"",""version"":""FinSearchComp Global subset""}"
finsearchcompt23,FinSearchComp T2&T3,Search Agent,%,,https://huggingface.co/moonshotai/Kimi-K2.5,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Kimi K2.5 model card."",""range"":[0,100],""tools"":""agentic"",""version"":""FinSearchComp T2&T3""}"
flenqa_3k,FlenQA (3K-token),Long Context,,,https://arxiv.org/abs/2402.14848,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Long-context QA at 3K tokens."",""range"":[0,100],""version"":""FlenQA 3K-token subset""}"
fleurs,FLEURS,Audio,reported aggregate WER/CER (lower=better),,https://huggingface.co/datasets/google/fleurs,"{""higher_is_better"":false,""metric_type"":""wer"",""multimodal_input"":true,""notes"":""Locale scope varies by source and must be recorded per score. Gemma Table 7 uses a simple macro over 12 locales, with CER for Korean, Japanese and Chinese."",""range"":[0,1],""tools"":""none"",""version"":""FLEURS ASR reported language aggregate""}"
frames,Frames,Agentic search,%,824.0,https://arxiv.org/abs/2409.12941,"{""higher_is_better"":true,""judge"":""LLM judge/autorater"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""FRAMES (Factuality, Retrieval, And reasoning MEasurement Set) evaluates end-to-end RAG systems on 824 multi-hop questions requiring information from 2-15 Wikipedia articles. The official HF dataset google/frames-benchmark has one test split with 824 rows and provides prompt, gold answer, required Wikipedia links, and reasoning-type labels. The paper evaluates single-step settings (naive prompt, BM25-retrieved prompt, oracle prompt) and a multi-step retrieval pipeline where the model generates search queries, retrieves Wikipedia documents, and answers after iterative retrieval. Answers are free-form, so scoring uses an LLM autorater to check whether the candidate answer matches the gold answer; the paper reports 0.96 accuracy and Cohen's kappa 0.889 against human annotations for Gemini-Pro-1.5-0514 as autorating LLM. Kimi K2 Thinking reports Frames under its Agentic Search section; this metadata uses the official FRAMES paper/dataset for count and protocol."",""range"":[0,100],""sampling"":""single answer per question; multi-step retrieval variants iteratively generate search queries"",""tools"":""retrieval/search tools over Wikipedia; official baselines include naive prompting, BM25 retrieval, oracle retrieval, and multi-step retrieval"",""version"":""FRAMES test set""}"
frontier_science_research,FrontierScience-Research,Science,reported score (%),,https://arxiv.org/abs/2601.21165,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official FrontierScience Research track; exact scored task count pending."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""research-oriented environment"",""version"":""FrontierScience-Research""}"
frontiermath,FrontierMath,Math,% correct T1-3,300.0,https://epoch.ai/benchmarks/frontiermath,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""tools"":""none"",""version"":""FrontierMath Tier 1-3 (300)""}"
frontiermath_tier4,FrontierMath Tier 4,Math,%,48.0,https://epoch.ai/benchmarks/frontiermath,"{""higher_is_better"":true,""judge"":""rule-based answer-function grader"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""FrontierMath Tier 4 is the hardest tier of FrontierMath. Epoch\u2019s official FrontierMath page documents the current Inspect-based evaluation: each question asks the model to solve a challenging mathematics problem, may use a Python tool, and must submit a Python function answer() through submit_answer. Correct answers receive 1 point and incorrect/no-submission answers receive 0. The answer function is executed with a 30-second runtime limit on typical 2025 commodity hardware. The page reports FrontierMath-Tier-4-2025-02-28-Private and FrontierMath-Tier-4-2025-07-01-Private evaluations with 48 samples/problems (e.g. Gemini 3 Pro: 3/48 API failures; Grok 4: 8/48 API errors). Use 48 as the scored Tier-4 item count. This protocol differs from OpenAI internal FrontierMath evaluations; OpenAI score pages remain score sources, while Epoch is the benchmark-definition source."",""range"":[0,100],""sampling"":""single evaluated run per problem; 1,000,000-token hard limit with forced submission after 660,000 tokens"",""tools"":""Python tool and submit_answer tool; code execution allowed during reasoning and answer grading"",""version"":""FrontierMath Tier 4 private set (48 problems)""}"
frontiersci_olympiad,FrontierScience-Olympiad,Science,reported score (%),,https://arxiv.org/abs/2601.21165,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official FrontierScience Olympiad track; exact scored task count pending."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""FrontierScience-Olympiad""}"
frontiersci_research,FrontierSci-research,Science,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""FrontierSci-research""}"
fsc_147,FSC-147 (lower=better),Vision Counting,metric (lower=better),,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":false,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":null,""tools"":""none"",""version"":""FSC-147 (lower=better)""}"
gaia,GAIA (text only),Agentic,%,103.0,https://arxiv.org/abs/2509.06501,"{""higher_is_better"":true,""judge"":""LLM-as-Judge for WebExplorer-style reported scores; GAIA original answers are unambiguous final-answer tasks"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""GAIA is a benchmark for general AI assistants with 466 total questions, requiring reasoning, tool use, web browsing, and sometimes multimodality. This BenchPress row is specifically GAIA (text only), not full GAIA. MiniMax-M2 reports GAIA (text only) using the same agent framework as WebExplorer and states that it uses the 103-sample text-only GAIA validation subset following WebExplorer. WebExplorer describes the GAIA setting as a widely adopted benchmark for General AI Assistants and uses a search/browse web-agent scaffold; it reports scores on information-seeking benchmarks using LLM-as-Judge, while the original GAIA benchmark defines unambiguous final-answer questions. Use 103 as the scored item count for this text-only subset."",""range"":[0,100],""sampling"":""single reported run not specified in MiniMax-M2 card; WebExplorer reports its own benchmark scores as Avg@4"",""tools"":""agentic web-search and browse tools"",""version"":""GAIA 103-sample text-only validation subset""}"
gdpval_aa_elo,GDPval (Artificial Analysis ELO),Knowledge,score,220.0,https://huggingface.co/datasets/openai/gdpval,"{""higher_is_better"":true,""judge"":""rubric-based grader / pairwise Elo aggregation in Artificial Analysis"",""metric_type"":""index"",""multimodal_input"":false,""notes"":""GDPval evaluates AI model performance on real-world economically valuable tasks. The official OpenAI HF dataset reports 220 tasks across 44 occupations; each task consists of a text prompt and supporting reference files, with expected deliverable files such as Excel workbooks, Word documents, PDFs, or other work products. Rows include human-authored rubric criteria with point values. This BenchPress row is the Artificial Analysis GDPval Elo/index view, so the score source is Artificial Analysis, but the benchmark-definition source and item count are the official OpenAI GDPval dataset. Use 220 as the scored task count; do not use the Knowledge category fallback."",""range"":null,""sampling"":""single deliverable per task; AA reports Elo-style score rather than raw percent"",""tools"":""office/document/spreadsheet workflow; reference files and deliverable files vary by task"",""version"":""GDPval public 220-task set; Artificial Analysis Elo aggregation""}"
gdpval_diamond,GDPVal-Diamond,Economic,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""GDPVal-Diamond""}"
gdpval_oai_woe,GDPval (OpenAI wins-or-ties),Office,%,,https://openai.com/index/introducing-gpt-5-5/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""OpenAI GDPval head-to-head wins or ties %, baseline against reference. Single-source: OpenAI GPT-5.5 blog."",""range"":[0,100],""tools"":""agentic"",""version"":""GDPval (OpenAI wins-or-ties)""}"
genebench,GeneBench,Science,%,,https://openai.com/index/introducing-gpt-5-5/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Genomics benchmark."",""range"":[0,100],""tools"":""none"",""version"":""GeneBench""}"
global_mmlu_lite,Global MMLU Lite,Knowledge,,7200.0,https://huggingface.co/datasets/CohereForAI/Global-MMLU-Lite,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Multilingual MMLU subset by Cohere. HF dataset card reports 18 languages with 400 test examples each. Distinct from MMLU (5-shot 14k EN) and MMMLU (multilingual MMLU full)."",""range"":[0,100],""version"":""Global MMLU Lite (multilingual MMLU subset, Cohere)""}"
global_piqa,Global PIQA,Reasoning,,6283.0,https://huggingface.co/datasets/mrlbenchmarks/global-piqa-parallel,"{""higher_is_better"":true,""judge"":""exact-match multiple-choice answer key"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Global PIQA is a participatory commonsense-reasoning benchmark for 100+ languages and cultures. The arXiv preprint describes 116 language varieties constructed by 335 researchers from 65 countries. The HF Global PIQA Parallel dataset card says each example has a question prompt and four candidate solutions, one correct and three incorrect; evaluation can use either prompted multiple-choice selection or completion likelihood ranking. The HF dataset-server size endpoint reports 6,283 total rows across the parallel test configurations, with 103 examples in each language-variety config that is populated in the dataset server. Google Gemini reports Global PIQA as commonsense reasoning across 100 languages and cultures; use the HF dataset row count as the scored item count."",""range"":[0,100],""sampling"":""single answer per multiple-choice question; prompted format selects A/B/C/D or completion format ranks candidate likelihoods"",""tools"":""none"",""version"":""Global PIQA parallel test set""}"
gpqa_diamond,GPQA Diamond,Science,multiple-choice accuracy (%),198.0,https://huggingface.co/datasets/Idavidrein/gpqa/tree/633f5ee89ab8ad4522a9f850766b73f62147ffdd,"{""harness"":""official GPQA"",""higher_is_better"":true,""judge"":""multiple-choice answer-key grading"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The pinned official dataset exposes 198 Diamond items. Repeated sampling and quantized decoding remain score-level settings."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""GPQA Diamond official 198-item split""}"
gpqa_main,GPQA Main (full set),Science,,448.0,https://arxiv.org/abs/2311.12022,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Full GPQA. Distinct from gpqa_diamond (198 hardest subset)."",""range"":[0,100],""version"":""GPQA full set (Rein et al. 2023, 448 questions)""}"
graphwalks_bfs_0k_128k,GraphWalks BFS 0-128K,Long Context,%,300.0,https://huggingface.co/datasets/openai/graphwalks,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""OpenAI GraphWalks is a multi-hop reasoning long-context benchmark. Each prompt gives a directed graph as an edge list and asks for either a BFS result set or a parent-node result set. The official HF dataset has 1,150 rows total, with columns prompt, answer_nodes, prompt_chars, problem_type, and date_added. Counting the official parquet files gives 550 BFS rows total; filtering to problem_type=bfs and prompt_chars<=128000 gives 300 rows for this BenchPress row. Outputs are parsed from a final \""Final Answer: [...]\"" list and scored by precision/recall/F1 set overlap against the answer nodes."",""range"":[0,100],""sampling"":""single response per graph operation; prompt includes 3-shot examples plus directed edge list and BFS operation"",""tools"":""none"",""version"":""GraphWalks BFS prompts with prompt_chars <=128K""}"
graphwalks_bfs_128k_plus,GraphWalks BFS >128k,Long Context,%,,https://openai.com/index/gpt-4-1/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per OpenAI GPT-4.1 blog."",""range"":[0,100],""tools"":""none"",""version"":""GraphWalks BFS >128k""}"
graphwalks_bfs_256k_1m,GraphWalks BFS 256K-1M,Long Context,% f1 (avg 256K-1M),,https://openai.com/index/introducing-gpt-5-5/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""tools"":""none"",""version"":""GraphWalks BFS 256K-1M""}"
graphwalks_parents_0k_128k,GraphWalks parents 0-128K,Long Context,%,350.0,https://huggingface.co/datasets/openai/graphwalks,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""OpenAI GraphWalks is a multi-hop reasoning long-context benchmark. Each prompt gives a directed graph as an edge list and asks for either a BFS result set or a parent-node result set. The official HF dataset has 1,150 rows total, with columns prompt, answer_nodes, prompt_chars, problem_type, and date_added. Counting the official parquet files gives 600 parent-node rows total; filtering to problem_type=parents and prompt_chars<=128000 gives 350 rows for this BenchPress row. Outputs are parsed from a final \""Final Answer: [...]\"" list and scored by precision/recall/F1 set overlap against the answer nodes."",""range"":[0,100],""sampling"":""single response per graph operation; prompt includes 3-shot examples plus directed edge list and parent-node operation"",""tools"":""none"",""version"":""GraphWalks parent-node prompts with prompt_chars <=128K""}"
graphwalks_parents_128k_plus,GraphWalks parents >128k,Long Context,%,,https://openai.com/index/gpt-4-1/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per OpenAI GPT-4.1 blog."",""range"":[0,100],""tools"":""none"",""version"":""GraphWalks parents >128k""}"
graphwalks_parents_256k_1m,GraphWalks parents 256K-1M,Long Context,%,,https://openai.com/index/introducing-gpt-5-4/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per OpenAI GPT-5.4 blog."",""range"":[0,100],""tools"":""none"",""version"":""GraphWalks parents 256K-1M""}"
gsm8k,GSM8K,Math,% correct,1319.0,https://arxiv.org/abs/2110.14168,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""tools"":""none"",""version"":""GSM8K (test, 1319 problems)""}"
hallusionbench,HallusionBench,Vision VQA,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""HallusionBench""}"
healthbench,HealthBench,Knowledge,%,5000.0,https://huggingface.co/datasets/openai/healthbench,"{""higher_is_better"":true,""judge"":""LLM judge over physician-written rubric criteria"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""HealthBench evaluates model responses to health and medical conversation prompts. The official OpenAI HF repository points to the HealthBench eval and OpenAI simple-evals reference implementation. The main OSS eval file 2025-05-07-06-14-12_oss_eval.jsonl has 5,000 rows. Each row contains a prompt conversation, example tags, and physician-written rubric criteria with point values; the public preview shows rubrics and prompt_id/canary fields. Separate official files exist for HealthBench Consensus (3,671 rows) and HealthBench Hard (1,000 rows), but this BenchPress row is the full HealthBench score, so use 5,000 items."",""range"":[0,100],""sampling"":""single assistant completion per medical conversation prompt; scored against rubric items with point values"",""tools"":""none"",""version"":""HealthBench full OSS eval set""}"
healthbench_consensus,HealthBench Consensus,Health,,,https://arxiv.org/abs/2508.10925,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Consensus subset of HealthBench."",""range"":[0,100],""version"":""HealthBench Consensus""}"
healthbench_hard,HealthBench Hard,Health,,,https://arxiv.org/abs/2508.10925,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Harder subset of HealthBench."",""range"":[0,100],""version"":""HealthBench Hard subset""}"
hellaswag,HellaSwag,Reasoning,% normalized multiple-choice accuracy,10042.0,https://huggingface.co/datasets/Rowan/hellaswag/resolve/218ec52e09a7e7462a5400043bb9a69a41d06b76/README.md,"{""higher_is_better"":true,""judge"":""length-normalized answer-choice accuracy"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official immutable dataset card reports 10,042 validation examples. The Xiaomi observation reports 10-shot prompting."",""range"":[0,100],""sampling"":""one four-choice response per validation example"",""tools"":""none"",""version"":""HellaSwag validation split""}"
hiddenmath,HiddenMath,Math,,,https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-2-0-Flash-Model-Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Google internal held-out math benchmark, AIME/AMC-style, not leaked online."",""range"":[0,100],""version"":""HiddenMath (held-out AIME/AMC-like, Google internal)""}"
hipho,HiPhO,Vision STEM,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""HiPhO""}"
hle,HLE (Humanity's Last Exam),Reasoning,% correct,2500.0,https://lastexam.ai/,"{""higher_is_better"":true,""judge"":""official answer-key rubric with o3-mini judge for non-exact answers"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Full finalized HLE contains 2,500 text and multimodal questions. Canonical setting is no tools; tool-enabled observations use the existing hle_tools row. Text-only observations use hle_text."",""range"":[0,100],""tools"":""none"",""version"":""Humanity's Last Exam (2500)""}"
hle_text,HLE Text,Reasoning,%,2158.0,https://labs.scale.com/leaderboard/humanitys_last_exam_text_only,"{""higher_is_better"":true,""judge"":""answer-key scoring for closed-ended answers"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Scale Labs states the text-only leaderboard evaluates text-based HLE questions excluding multimodal content and represents 86% of the finalized 2,500-question HLE. Public text-only mirror DongfuJiang/hle_text_only reports 2,158 rows, consistent with that 86% subset. HLE consists of multiple-choice and short-answer questions with unambiguous answers suitable for automated grading; Doubao Seed is retained only as a score source, not as the benchmark-definition source."",""range"":[0,100],""tools"":""none"",""version"":""Humanity's Last Exam text-only subset of finalized 2,500-question HLE""}"
hle_tools,HLE (w/ tools),Reasoning & Knowledge,accuracy (%),2500.0,https://raw.githubusercontent.com/centerforaisafety/hle/73ae974b1844c3ffa64c3f4343d9f1f259575700/README.md,"{""harness"":""source-specific tool harness"",""higher_is_better"":true,""judge"":""official answer-key and HLE grading protocol"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""The pinned official HLE repository defines the finalized 2,500-question release. Tool suites and context management remain score-level settings."",""range"":[0,100],""sampling"":""one scored answer per question"",""tools"":""source-reported search/code/web tools"",""version"":""HLE finalized full 2,500-question text+image set""}"
hle_verified,HLE Verified,Knowledge,accuracy (%),1811.0,https://raw.githubusercontent.com/SKYLENAGE-AI/HLE-Verified/b705e0fb541c025a1532ce0d60d70ae2f53b00e0/README.md,"{""dataset_split"":""Gold + Revision"",""harness"":""official/self-computed"",""higher_is_better"":true,""judge"":""HLE-Verified answer evaluator"",""metric_type"":""accuracy_pct"",""multimodal_input"":true,""notes"":""Google explicitly reports accuracy over all 1,811 verified/revised items and excludes the 689 Uncertain items."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""HLE-Verified full verified set: Gold 668 + Revision 1,143; Uncertain 689 excluded""}"
hle_vl,HLE-VL,Vision Agent,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""search"",""version"":""HLE-VL""}"
hmmt_feb_2025,HMMT Feb 2025,Math,%,30.0,https://huggingface.co/datasets/MathArena/hmmt_feb_2025,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""MathArena dataset has 30 questions; MathArena evaluates each model 4 times per problem."",""range"":[0,100],""sampling"":""samples=4"",""tools"":""none"",""version"":""HMMT Feb 2025""}"
hmmt_feb_2026,HMMT Feb 2026,Math,% correct (pass@1),33.0,https://huggingface.co/datasets/MathArena/hmmt_feb_2026,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""sampling"":""samples=4"",""tools"":""none"",""version"":""HMMT Feb 2026""}"
hmmt_nov_2025,HMMT Nov 2025,Math,% correct,30.0,https://huggingface.co/datasets/MathArena/hmmt_nov_2025,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""sampling"":""samples=4"",""tools"":""none"",""version"":""HMMT Nov 2025""}"
humaneval,HumanEval,Coding,pass@1 %,164.0,https://github.com/openai/human-eval,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""tools"":""none"",""version"":""HumanEval (Chen et al. 2021)""}"
humaneval_plus,HumanEval+,Coding,pass@1 (%),164.0,https://huggingface.co/datasets/evalplus/humanevalplus/resolve/d32357cf319e50e9c8d8dab5ea876c72b0fd321b/README.md,"{""higher_is_better"":true,""judge"":""EvalPlus expanded unit-test execution"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official immutable dataset card reports 164 problems."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""code execution for tests"",""version"":""EvalPlus HumanEval+ immutable test split""}"
ib_modeling,Investment Banking Modeling Tasks (Internal),Finance,%,,https://openai.com/index/introducing-gpt-5-5/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Internal OpenAI IB modeling benchmark."",""range"":[0,100],""tools"":""none"",""version"":""Investment Banking Modeling Tasks (Internal)""}"
ifbench,IFBench,Instruction Following,% correct,300.0,https://github.com/allenai/IFBench,"{""higher_is_better"":true,""judge"":""rule-based verification functions"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Correct benchmark source is AllenAI IFBench / arXiv 2507.02833, not the previously listed arXiv 2502.09980 V2V-QA paper. IFBench has 58 out-of-domain verifiable constraints; the final single-turn benchmark has 300 prompts, matching the allenai/IFBench_test HF dataset size endpoint. The paper generally reports prompt-level loose accuracy with automatic verifier functions."",""range"":[0,100],""tools"":""none"",""version"":""IFBench single-turn test set""}"
ifeval,IFEval,Instruction Following,% correct (prompt strict),541.0,https://arxiv.org/abs/2311.07911,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""tools"":""none"",""version"":""IFEval prompt-strict (541)""}"
imo_2025,IMO 2025,Math,% of 42 points,6.0,https://matharena.ai/imo/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""tools"":""none"",""version"":""IMO 2025""}"
imo_answerbench,IMO-AnswerBench,Math,,400.0,https://imobench.github.io/,"{""higher_is_better"":true,""judge"":""Gemini 2.5 Pro AnswerAutoGrader"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""IMO-Bench official site and arXiv 2511.01846 define IMO-AnswerBench as 400 Olympiad short-answer problems. The paper uses AnswerAutoGrader, built with Gemini 2.5 Pro, to extract final answers and assess correctness against ground truth. Kimi K2.5 reports IMO-AnswerBench with avg@8."",""range"":[0,100],""sampling"":""samples=8 (reported as avg@8)"",""version"":""IMO-AnswerBench""}"
infobench,InFoBench,Instruction following,,2250.0,https://github.com/qinyiwei/InfoBench,"{""higher_is_better"":true,""judge"":""GPT-4-0314 judge for decomposed yes/no requirement questions"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official InfoBench repo, arXiv 2401.03601, and HF kqsong/InFoBench define 500 instructions and 2,250 decomposed questions. The DRFR metric scores whether each decomposed requirement is satisfied, and the official evaluation script uses GPT-4-0314 by default to answer each decomposed yes/no question at temperature 0. Cohere Command A remains only a score source for existing cells."",""range"":[0,100],""sampling"":""greedy decoding; judge temperature=0"",""tools"":""none"",""version"":""InFoBench: 500 instructions with 2,250 decomposed requirement-level scoring units""}"
infovqa,InfoVQA (val),Vision,%,,https://huggingface.co/moonshotai/Kimi-K2.5,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Kimi K2.5 model card."",""range"":[0,100],""tools"":""none"",""version"":""InfoVQA (val)""}"
internal_api_if_hard,Internal API IF Hard,Instruction Following,%,,https://openai.com/index/introducing-gpt-5-for-developers/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""OpenAI GPT-5 developer blog states that the internal OpenAI API instruction-following eval uses difficult instructions derived from real developer feedback and that reasoning models were run with high reasoning effort. The GPT-4.1 API blog describes the same internal instruction-following eval as covering format following, negative instructions, ordered instructions, content requirements, ranking, and overconfidence, split into easy, medium, and hard prompts. OpenAI does not disclose item count or scoring implementation, so keep num_problems null rather than converting the 500 category fallback into a source-backed count."",""range"":[0,100],""sampling"":""pass@1; reasoning models run with high reasoning effort"",""tools"":""none"",""version"":""OpenAI internal API instruction-following eval, hard prompts""}"
inverse_ifeval,Inverse IFEval,Instruction Following,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""Inverse IFEval""}"
korbench,KORBench,Reasoning,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""KORBench""}"
livebench,LiveBench,Composite,overall score,1000.0,https://github.com/LiveBench/LiveBench,"{""higher_is_better"":true,""judge"":""objective ground-truth scoring without LLM evaluators"",""metric_type"":""index"",""multimodal_input"":false,""notes"":""Official LiveBench README defines 18 tasks across 6 categories and states that each question has verifiable objective ground-truth answers, scored automatically without an LLM judge. The README says the current 2025-04-25 release is not fully public on Hugging Face and recommends --livebench-release-option 2024-11-25 for the most recent public full-category evaluation. Applying the official HF release/removal filter to the six livebench category datasets gives 1,000 active questions for 2024-11-25: coding 128, data_analysis 150, instruction_following 200, math 232, reasoning 150, language 140. Later public HF rows are incomplete for full-category evaluation."",""range"":null,""sampling"":""pass@1"",""tools"":""none"",""version"":""LiveBench 2024-11-25 full public release""}"
livecodebench,LiveCodeBench,Coding,pass@1 %,1055.0,https://livecodebench.github.io/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=agentic (scaffold-defined). Preferred = official harness tools (bash/editor for SWE-bench, browser for OSWorld/BrowseComp, official APIs for tau-bench/MCPAtlas/Toolathlon, terminal for Terminal-Bench). Non-official scaffolds \u2192 matches_canonical=false."",""range"":[0,100],""tools"":""agentic"",""version"":""LiveCodeBench (1055)""}"
livecodebench_pro,LiveCodeBench Pro (Elo),Coding,,,https://livecodebench.github.io/pro.html,"{""higher_is_better"":true,""metric_type"":""elo"",""multimodal_input"":false,""notes"":""Elo rating against competitive programming pool."",""range"":[0,4000],""version"":""LiveCodeBench Pro \u2014 Codeforces/ICPC/IOI competitive set""}"
livecodebench_v5,LiveCodeBench v5,Coding,%,,https://arxiv.org/abs/2504.13914,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Seed-Thinking-v1.5 paper."",""range"":[0,100],""tools"":""none"",""version"":""LiveCodeBench v5""}"
livecodebench_v6,LiveCodeBench v6,Coding,pass@1 (%),1055.0,https://github.com/LiveCodeBench/LiveCodeBench,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""1,055 code-generation problems from May 2023 through April 2025; generated code is executed by the judge."",""range"":[0,100],""tools"":""none"",""version"":""LiveCodeBench release_v6""}"
livesports_3k,LiveSports-3K,Video,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""LiveSports-3K""}"
loft_128k,LOFT (128k),Long Context,%,,https://x.ai/news/grok-3,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per xAI Grok 3 blog."",""range"":[0,100],""tools"":""none"",""version"":""LOFT (128k)""}"
logicvista,LogicVista,Vision Reasoning,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""LogicVista""}"
longbench_v2,LongBench-V2,Long Context,%,503.0,https://huggingface.co/datasets/THUDM/LongBench-v2,"{""higher_is_better"":true,""judge"":""exact multiple-choice answer matching"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official LongBench v2 sources are the THUDM LongBench repo, HF dataset, and arXiv:2412.15204, not the prior DeepSeek model card. The paper/README/HF card state 503 challenging multiple-choice questions with contexts from 8k to 2M words across six task categories. Loading THUDM/LongBench-v2 split=train returns 503 rows with answer keys A-D. The official pred.py default runs one direct-answer generation per item; --cot, --rag, and --no_context are optional settings. The official result.py reports overall accuracy as exact match between extracted option and the answer."",""range"":[0,100],""sampling"":""pass@1; temperature=0.1 in official script"",""tools"":""none"",""version"":""LongBench-V2""}"
longdocurl,LongDocURL,Vision Long Context,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""LongDocURL""}"
longfact_concepts,LongFact-Concepts (hallucination rate),Hallucination,%,1140.0,https://github.com/google-deepmind/long-form-factuality/tree/main/longfact,"{""higher_is_better"":false,""judge"":""SAFE LLM-as-a-judge factuality evaluator"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official LongFact sources are the google-deepmind/long-form-factuality repo and arXiv:2403.18802, not the prior OpenAI GPT-5 model blog. The longfact README states that LongFact-Concepts has the same 38 topics as LongFact-Objects and 30 unique prompts per topic, giving 1,140 prompts for the Concepts subtask and 2,280 prompts for the full LongFact benchmark. The SAFE README describes evaluation as an LLM-based pipeline that decomposes each long-form response into atomic facts, revises facts to be self-contained, classifies relevance, and checks support using Google Search calls; reported hallucination rate is therefore search-augmented posthoc factuality scoring. LOWER IS BETTER."",""range"":[0,100],""sampling"":""pass@1 model response; SAFE max_steps=5 and num_searches=3 by default"",""tools"":""Google Search via Serper in SAFE evaluation"",""version"":""LongFact-Concepts (hallucination rate)""}"
longfact_objects,LongFact-Objects (hallucination rate),Hallucination,%,1140.0,https://github.com/google-deepmind/long-form-factuality/tree/main/longfact,"{""higher_is_better"":false,""judge"":""SAFE LLM-as-a-judge factuality evaluator"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official LongFact sources are the google-deepmind/long-form-factuality repo and arXiv:2403.18802, not the prior OpenAI GPT-5 model blog. The longfact README states that LongFact-Objects has the same 38 topics as LongFact-Concepts and 30 unique prompts per topic, giving 1,140 prompts for the Objects main task and 2,280 prompts for the full LongFact benchmark. The SAFE README describes evaluation as an LLM-based pipeline that decomposes each long-form response into atomic facts, revises facts to be self-contained, classifies relevance, and checks support using Google Search calls; reported hallucination rate is therefore search-augmented posthoc factuality scoring. LOWER IS BETTER."",""range"":[0,100],""sampling"":""pass@1 model response; SAFE max_steps=5 and num_searches=3 by default"",""tools"":""Google Search via Serper in SAFE evaluation"",""version"":""LongFact-Objects (hallucination rate)""}"
longform_writing,Longform Writing,Writing,%,,https://huggingface.co/moonshotai/Kimi-K2-Thinking,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Kimi K2-Thinking model card."",""range"":[0,100],""tools"":""none"",""version"":""Longform Writing""}"
longvideobench,LongVideoBench,Vision,%,,https://huggingface.co/moonshotai/Kimi-K2.5,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Kimi K2.5 model card."",""range"":[0,100],""tools"":""none"",""version"":""LongVideoBench""}"
lpfqa,LPFQA,Knowledge,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""LPFQA""}"
lvbench,LVBench,Multimodal,multiple-choice accuracy (%),1549.0,https://raw.githubusercontent.com/zai-org/LVBench/518df47219862534dad39fa1373b4e7c862a4cd5/README.md,"{""higher_is_better"":true,""metric_type"":""accuracy_pct"",""multimodal_input"":true,""notes"":""Official LVBench release: 103 videos and 1,549 multiple-choice question-answer pairs."",""range"":[0,100],""tools"":""none"",""version"":""LVBench extreme long-video benchmark""}"
mars_bench,MARS-Bench,Instruction Following,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""MARS-Bench""}"
math,MATH,Math,% exact-answer accuracy,5000.0,https://arxiv.org/pdf/2103.03874v2,"{""higher_is_better"":true,""judge"":""normalized final-answer exact match"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The official paper defines 12,500 total problems: 7,500 train and 5,000 test. Count is the 5,000 evaluated test problems, not the full corpus."",""range"":[0,100],""sampling"":""one response per test problem"",""tools"":""none"",""version"":""MATH competition benchmark test split""}"
math_500,MATH-500,Math,% correct,500.0,https://arxiv.org/abs/2103.03874,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""tools"":""none"",""version"":""MATH-500 subset (Hendrycks)""}"
matharena_apex_2025,MathArena Apex 2025,Math,% correct,12.0,https://matharena.ai/apex/,"{""higher_is_better"":true,""judge"":""exact final-answer auto-verification"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official MathArena Apex page defines the current benchmark as 12 hard final-answer problems selected from 2025 competitions. Problem filtering used 4 attempts with frontier models, but the reported aggregate results evaluate 9 models with 16 independent runs per problem and report the average success rate over all problems and attempts. GPT-5 with scaffolding is a separate elicitation condition using 4 runs and should not define the canonical direct-prompt setting. Tools are none for the direct reasoning models; final answers are automatically verified."",""range"":[0,100],""sampling"":""samples=16 independent runs per problem in the main Apex table"",""tools"":""none"",""version"":""MathArena Apex 2025""}"
mathcanvas,MathCanvas,Math,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""MathCanvas""}"
mathkangaroo,MathKangaroo,Math,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""MathKangaroo""}"
mathvision,MathVision,Math,% correct,3040.0,https://huggingface.co/datasets/MathLLMs/MathVision,"{""higher_is_better"":true,""judge"":""rule-based answer extraction and symbolic/option matching"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Official MATH-Vision sources are the mathllm/MATH-V repo, MathLLMs/MathVision HF dataset, project page, and arXiv:2402.14804. The paper/project page define MATH-Vision as 3,040 mathematical problems with visual contexts across 16 subjects and 5 difficulty levels. The HF dataset has test=3,040 and testmini=304; the project page main leaderboard is on the full 3,040-example test set, while testmini is used for human performance and some smaller evaluations. Loading MathLLMs/MathVision confirms test has 3,040 rows and testmini has 304 rows. Official evaluation/evaluate.py computes overall accuracy by extracting a model answer and checking it against the answer key or option text with symbolic/equivalence rules. VLMEvalKit later supports LLM-based answer extraction, but this is extraction, not LLM-as-a-judge scoring."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""MATH-Vision full test split""}"
mathvista,MathVista,Math/Vision,%,1000.0,https://huggingface.co/datasets/AI4Math/MathVista,"{""higher_is_better"":true,""judge"":""rule-based answer extraction and answer-key matching"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Official MathVista sources are the project page, AI4Math/MathVista HF dataset, lupantech/MathVista repo, and arXiv:2310.02255, not the prior OpenAI GPT-4.1 model blog. The paper/project page define the full dataset as 6,141 examples from 31 datasets. HF splits are testmini=1,000 with public answer labels and test=5,141 for private standard evaluation; the HF card says the available leaderboard is testmini, while test labels are not public. Because the OpenAI GPT-4.1 source table reports only 'MathVista' without specifying private test, this canonical setting uses the public testmini subset. Evaluation uses image+question inputs; outputs are scored by extracting an answer and matching the answer key, not by LLM-as-a-judge."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""MathVista public testmini subset""}"
mathvista_mini,MathVista (mini),Multimodal Math,,,https://huggingface.co/moonshotai/Kimi-K2.5,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""avg@3."",""range"":[0,100],""version"":""MathVista mini""}"
mbpp_plus,MBPP+,Coding,pass@1 (%),378.0,https://huggingface.co/datasets/evalplus/mbppplus/resolve/b2d74c91837c3f2a20c1299ae98133cbe7cfa077/README.md,"{""higher_is_better"":true,""judge"":""EvalPlus expanded unit-test execution"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official immutable dataset card reports 378 problems."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""code execution for tests"",""version"":""EvalPlus MBPP+ immutable test split""}"
mcpatlas,MCPAtlas Public,Agentic,% correct (pass@1),500.0,https://huggingface.co/datasets/ScaleAI/MCP-Atlas,"{""higher_is_better"":true,""judge"":""Gemini 2.5 Pro claims-based coverage judge; pass if coverage >= 0.75"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""May 2026 public release contains 500 tasks, 36 MCP servers and 220 tools, scored by claim coverage >=0.75. The current repo later expanded to 307 tools and a different default judge/turn cap; score-level settings must preserve the historical Scale leaderboard snapshot."",""range"":[0,100],""sampling"":""pass@1; public harness default maxTurns=20"",""tools"":""agentic MCP servers in the official containerized harness"",""version"":""MCP-Atlas public 500-task May 2026 leaderboard release""}"
mcpmark,MCPMark,Agentic,% success (pass@1),127.0,https://github.com/eval-sys/mcpmark,"{""higher_is_better"":true,""judge"":""programmatic verification scripts"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official sources are arXiv:2509.24002, mcpmark.ai, and the eval-sys/mcpmark GitHub repo. MCPMark standard contains 127 tasks with curated initial states and verify.py scripts: 30 Filesystem, 28 Notion, 23 GitHub, 21 PostgreSQL, and 25 Playwright/Playwright-WebArena tasks. GitHub tree count confirms 127 meta.json files under tasks/*/standard; the additional 50 easy tasks are a later lightweight smoke-test suite and are not part of the canonical standard benchmark. Evaluation uses MCPMark-Agent in a tool-calling loop, max 100 turns and 3600-second timeout, then checks final environment state with programmatic verification."",""range"":[0,100],""sampling"":""pass@1; trials=4 independent runs in the official paper pass@1 mean; pass@4 and pass^4 also reported"",""tools"":""agentic MCP tool-calling loop over Notion, GitHub, Filesystem, PostgreSQL, and Playwright/WebArena"",""version"":""MCPMark standard task suite""}"
medxpertqa_mm,MedXpertQA MM,Medical,multiple-choice accuracy (%),2000.0,https://github.com/TsinghuaC3I/MedXpertQA,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""2,000 multimodal medical multiple-choice test questions."",""range"":[0,100],""tools"":""none"",""version"":""MedXpertQA MM test split""}"
medxpertqa_text,MedXpertQA (Text),Medical,,,,"{""higher_is_better"":true,""judge"":""gpt-oss-120b"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Meta MSL eval methodology: 2,450 prompts spanning medical specialties; 10 answer choices; graded with gpt-oss-120b. Source: https://ai.meta.com/blog/introducing-muse-spark-msl/"",""range"":[0,100],""version"":""MedXpertQA Text (2,450 prompts, 10-choice A\u2013J)""}"
mgsm,MGSM,Math,exact match (%),2500.0,https://github.com/google-research/url-nlp/tree/main/mgsm,"{""higher_is_better"":true,""judge"":""rule-based exact match on numeric answer"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official source is google-research/url-nlp MGSM, introduced by arXiv:2210.03057. The benchmark manually translates the same 250 GSM8K test problems into 10 languages (Spanish, French, German, Russian, Chinese, Japanese, Thai, Swahili, Bengali, Telugu), giving 2,500 multilingual scored prompts. The official repo also includes an English TSV with 250 rows; harnesses such as OpenAI simple-evals may include English for 2,750 total prompts, but the MGSM benchmark definition is the 10-language translation set."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""MGSM 10-language benchmark""}"
minedojo_verified,Minedojo Verified,Vision Agent,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""env"",""version"":""Minedojo Verified""}"
minerva,MINERVA,Video,five-choice accuracy (%),1515.0,https://arxiv.org/abs/2505.00681,"{""harness"":""official"",""higher_is_better"":true,""judge"":""exact multiple-choice grader"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Original MINERVA, not Cultural or Ego variants."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""MINERVA""}"
mm_browsecomp,MM-BrowseComp,Vision Agent,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""search"",""version"":""MM-BrowseComp""}"
mme_cc,MME-CC,Vision VQA,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""MME-CC""}"
mmlongbench,MMLongBench-128K v1.1,Vision Long Context,aggregate score (%),13331.0,https://github.com/EdinburghNLP/MMLongBench,"{""harness"":""official"",""higher_is_better"":true,""judge"":""mixed rule-based and LLM-based task metrics"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""v1.1 128K evaluation across five task families."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""MMLongBench-128K v1.1""}"
mmlongbench_doc,MMLongBench-Doc,Vision Long Context,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""MMLongBench-Doc""}"
mmlu,MMLU,Knowledge,% correct,14042.0,https://arxiv.org/abs/2009.03300,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""tools"":""none"",""version"":""MMLU (5-shot, 14042 questions)""}"
mmlu_pro,MMLU-Pro,Knowledge,% correct,12032.0,https://arxiv.org/abs/2406.01574,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""tools"":""none"",""version"":""MMLU-Pro""}"
mmlu_redux,MMLU-Redux,Knowledge,% accuracy,3000.0,https://huggingface.co/datasets/edinburgh-dawg/mmlu-redux/resolve/3720db6aeb3d019de48bf37916c1a54074ff4997/README.md,"{""higher_is_better"":true,""judge"":""answer-key accuracy"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The immutable original MMLU-Redux release has 30 subjects with 100 test questions each, totaling 3,000. It is distinct from MMLU-Redux 2.0."",""range"":[0,100],""sampling"":""one multiple-choice response per question"",""tools"":""none"",""version"":""MMLU-Redux original 30-subject test release""}"
mmmlu,MMMLU,Knowledge,% correct,196588.0,https://huggingface.co/datasets/openai/MMMLU,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""14 locales times 14,042 MMLU test questions = 196,588 scored prompts."",""range"":[0,100],""tools"":""none"",""version"":""MMMLU 14 translated MMLU test locales""}"
mmmu,MMMU,Multimodal,% correct,900.0,https://mmmu-benchmark.github.io/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""tools"":""none"",""version"":""MMMU validation (900 questions)""}"
mmmu_pro,MMMU-Pro,Multimodal,% correct,3460.0,https://huggingface.co/datasets/MMMU/MMMU_Pro,"{""higher_is_better"":true,""judge"":""rule-based multiple-choice accuracy"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Official sources are arXiv:2409.02813 and the MMMU/MMMU_Pro HuggingFace dataset. The paper filters MMMU to 1,730 questions, augments them to the standard 10-option setting, and creates a matching vision-only version; official overall MMMU-Pro is the average of standard (10 options) and vision, so the evaluated prompt count is 1,730 + 1,730 = 3,460. The standard (4 options) split is a comparison setting and is not counted in the canonical overall. tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""MMMU-Pro overall: standard (10 options) + vision""}"
mmsibench_circular,MMSIBench (circular),Vision Spatial,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""MMSIBench (circular)""}"
mmstar,MMStar,Vision VQA,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""MMStar""}"
mmvu,MMVU,Vision,%,,https://huggingface.co/moonshotai/Kimi-K2.5,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Kimi K2.5 model card."",""range"":[0,100],""tools"":""none"",""version"":""MMVU""}"
morse_500,Morse-500,Video,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""Morse-500""}"
motionbench,MotionBench,Vision,%,,https://huggingface.co/moonshotai/Kimi-K2.5,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Kimi K2.5 model card."",""range"":[0,100],""tools"":""none"",""version"":""MotionBench""}"
mrcr_v1,MRCR v1,Long-context,,2000.0,https://storage.googleapis.com/deepmind-media/gemini/gemini_v1_5_report.pdf,"{""higher_is_better"":true,""judge"":""difflib SequenceMatcher string-similarity ratio"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official source is the Gemini 1.5 technical report. MRCR presents a long user-model conversation with two adversarially similar writing requests and asks the model to reproduce the response associated with a target request. Figure 12 reports cumulative average string-similarity score as a function of context length over 2,000 MRCR instances, up to 1M tokens. This v1 1M-context setting is distinct from OpenAI MRCR v2 128k / 2-needle / 8-needle variants."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""MRCR v1 at up to 1M context""}"
mrcr_v2,MRCR v2,Long Context,% correct,2400.0,https://huggingface.co/datasets/openai/mrcr,"{""higher_is_better"":true,""judge"":""difflib SequenceMatcher string-similarity ratio with required hash prefix"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official source is the openai/mrcr HuggingFace dataset. OpenAI MRCR v2 expands Gemini MRCR into an open long-context multiple-needle benchmark with 2, 4, or 8 identical asks hidden in a synthetic conversation. The dataset has 100 samples per bin, 8 token-length bins from 4k through 1M, and three needle settings, giving 3 * 8 * 100 = 2,400 rows. Many model cards report the 128k slice; the full released dataset also includes 262k, 524k, and 1M bins. tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""OpenAI MRCR v2 full dataset""}"
mrcr_v2_2needle_128k,"OpenAI MRCR v2 (2 needle, 128k)",Long Context,%,500.0,https://huggingface.co/datasets/openai/mrcr,"{""higher_is_better"":true,""judge"":""difflib SequenceMatcher string-similarity ratio with required hash prefix"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official source is the openai/mrcr HuggingFace dataset, with results reported in the OpenAI GPT-4.1 blog. This row is the 2-needle 128k slice: one needle setting, five token-length bins up to 131,072 tokens, and 100 samples per bin, giving 500 scored prompts. The full 2-needle dataset has 800 rows across all eight bins up to 1M."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""OpenAI MRCR v2 (2 needle, 128k)""}"
mrcr_v2_2needle_1m,"OpenAI MRCR v2 (2 needle, 1M)",Long Context,%,,https://openai.com/index/gpt-4-1/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per OpenAI GPT-4.1 blog."",""range"":[0,100],""tools"":""none"",""version"":""OpenAI MRCR v2 (2 needle, 1M)""}"
mrcr_v2_2needle_256k,"OpenAI MRCR v2 (2-needle, 256k)",Long Context,,,,"{""judge"":""rule-based"",""notes"":""Per OpenAI GPT-5 developer blog https://openai.com/index/introducing-gpt-5-for-developers/""}"
mrcr_v2_8needle,OpenAI MRCR v2 (8-needle),Long Context,%,800.0,https://huggingface.co/datasets/openai/mrcr,"{""higher_is_better"":true,""judge"":""difflib SequenceMatcher string-similarity ratio with required hash prefix"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official source is the openai/mrcr HuggingFace dataset. This row is the 8-needle slice: one needle setting, eight token-length bins from 4k through 1M, and 100 samples per bin, giving 800 scored prompts. Some model cards report restricted context slices, but the benchmark variant name here does not restrict to 128k."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""OpenAI MRCR v2 (8-needle)""}"
mt_aime_2024,MT-AIME2024,Math,%,1650.0,https://huggingface.co/datasets/amphora/MCLM,"{""higher_is_better"":true,""judge"":""rule-based verifier"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official source is the MCLM dataset and Son et al. (arXiv:2502.17407). MT-AIME2024 translates the full AIME 2024 set into 55 languages; the dataset stores 30 rows with 55 language columns, so the canonical evaluation contains 1,650 language-specific scored prompts."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""MCLM MT-AIME2024 (multilingual AIME 2024, 55 languages, Son et al. 2025)""}"
mt_bench_101,MT-Bench-101,Chat,Score (1-10),,https://github.com/InternLM/InternLM,"{""higher_is_better"":true,""metric_type"":""raw"",""multimodal_input"":false,""notes"":""Per InternLM3 GitHub README. MT-Bench-101 scored 1-10."",""range"":[1,10],""tools"":""none"",""version"":""MT-Bench-101 (Score 1-10)""}"
mtvqa,MTVQA,Vision VQA,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""MTVQA""}"
muirbench,MUIRBench,Vision VQA,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""MUIRBench""}"
multi_if,Multi-IF,Instruction Following,%,13503.0,https://huggingface.co/datasets/facebook/Multi-IF,"{""higher_is_better"":true,""judge"":""script-based verifiable-instruction checks"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official sources are the facebook/Multi-IF HuggingFace dataset and He et al. (arXiv:2410.15553). The dataset has 4,501 multilingual conversations across 8 languages, and each conversation has three turns; the cost proxy counts the 13,503 model-turn generations that must be evaluated. The reported metric averages instruction-level strict accuracy, conversation-level strict accuracy, instruction-level loose accuracy, and conversation-level loose accuracy across languages and turns."",""range"":[0,100],""sampling"":""single response per turn"",""tools"":""none"",""version"":""Multi-IF (8-language, 3-turn conversations)""}"
multi_swe_bench,Multi-SWE-bench,Coding,%,1632.0,https://huggingface.co/datasets/ByteDance-Seed/Multi-SWE-bench,"{""higher_is_better"":true,""judge"":""execution-based patch validation"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official sources are the ByteDance-Seed/Multi-SWE-bench HuggingFace dataset and Zan et al. (arXiv:2504.02605). The full benchmark covers Java, TypeScript, JavaScript, Go, Rust, C, and C++ with 1,632 human-validated issue-resolving instances curated from 2,456 candidates. Later auxiliary releases such as mini, flash, RL, and Python supplement files are not counted in this canonical full benchmark row."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""repository checkout plus Docker/unit-test execution"",""version"":""Multi-SWE-bench (full 7-language issue-resolving benchmark)""}"
multichallenge,MultiChallenge,Instruction Following,%,273.0,https://github.com/ekwinox117/multi-challenge,"{""higher_is_better"":true,""judge"":""automated LLM judge with instance-level rubrics"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official sources are the MultiChallenge paper (arXiv:2501.17399) and the released benchmark_questions.jsonl in the project repository. The benchmark contains 273 maximum-10-turn test conversations: 113 inference-memory, 69 instruction-retention, 41 reliable-version-editing, and 50 self-coherence conversations. Cost counts one model completion per conversation because each row provides conversation history ending in a target question."",""range"":[0,100],""sampling"":""single final response per conversation"",""tools"":""none"",""version"":""MultiChallenge""}"
multichallenge_o3mini_grader,MultiChallenge (o3-mini grader),Instruction Following,%,273.0,https://github.com/ekwinox117/multi-challenge,"{""higher_is_better"":true,""judge"":""o3-mini grader / LLM-as-judge with instance-level binary rubrics"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""MultiChallenge has 273 test conversations in the paper and official GitHub data. Each item requires one model response to a multi-turn conversation history, then an LLM grader evaluates the final response against an instance-level rubric. HF currently reports 266 rows, treated as a conflicting mirror/snapshot rather than the canonical paper/repo count."",""range"":[0,100],""sampling"":""attempts=1 unless otherwise reported"",""tools"":""none"",""version"":""Scale MultiChallenge official GitHub benchmark_questions.jsonl / paper Table 1""}"
multilingual_mmlu,Multilingual MMLU,Multilingual,,,https://huggingface.co/microsoft/Phi-4-mini-instruct,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""5-shot MMLU across multiple languages."",""range"":[0,100],""version"":""Multilingual MMLU (5-shot)""}"
multipl_e_avg,MultiPL-E (average),Coding,%,12667.0,https://huggingface.co/datasets/nuprl/MultiPL-E,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""MultiPL-E is a multilingual code-generation benchmark translated from HumanEval and MBPP. The HF dataset-server reports 12,667 total test rows across 47 configs (3,811 HumanEval rows and 8,856 MBPP rows). If the score source used only a HumanEval subset, 3,811 is the narrower count; either interpretation is Tier 3 under the cost proxy."",""range"":[0,100],""sampling"":""0-shot/pass@1 as reported by Mistral Medium 3 blog"",""tools"":""code execution"",""version"":""MultiPL-E full public HF dataset, averaged across language/config rows""}"
nl2repo_bench,NL2Repo-Bench,Repository Code,%,104.0,https://arxiv.org/abs/2512.12730,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""NL2Repo-Bench contains 104 repository-generation tasks. Each task gives a natural-language requirements document and empty workspace; generated repositories are evaluated with original upstream pytest suites. Item count is task instances, not upstream test cases."",""range"":[0,100],""sampling"":""Pass@1 unless otherwise reported"",""tools"":""agentic code execution"",""version"":""NL2Repo-Bench full benchmark""}"
nl2repo_pass1,NL2Repo (Pass@1),Repository Code,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""code execution"",""version"":""NL2Repo (Pass@1)""}"
ntrex,NTREX,Multilingual,,,https://cohere.com/research/papers/command-a-technical-report.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Machine translation COMET-20 score."",""range"":[0,100],""version"":""NTREX (COMET-20)""}"
ocrbench,OCRBench,Multimodal,,,https://huggingface.co/moonshotai/Kimi-K2.5,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""OCR benchmark."",""range"":[0,100],""version"":""OCRBench""}"
ocrbench_v2,OCRBench v2,Document/Chart,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""OCRBench v2""}"
odvbench,ODVBench,Video,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""ODVBench""}"
officeqa,OfficeQA,Office,exact-match accuracy (%),246.0,https://www.databricks.com/blog/introducing-officeqa-benchmark-end-to-end-grounded-reasoning,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""OfficeQA productivity benchmark (different from OfficeQA Pro). Per OpenAI GPT-5.4 blog."",""range"":[0,100],""tools"":""agentic"",""version"":""OfficeQA""}"
officeqa_pro,OfficeQA Pro,Office,exact-match accuracy (%),133.0,https://arxiv.org/abs/2603.08655,"{""harness"":""OfficeQA Pro official harness"",""higher_is_better"":true,""judge"":""OfficeQA Pro exact-match accuracy"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Every PDF is rendered as images; no machine-readable PDF text is supplied."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""document and rendered-image analysis"",""version"":""OfficeQA Pro 133-question release""}"
ojbench,OJBench,Coding,%,232.0,https://arxiv.org/abs/2506.16395,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""OJBench comprises 232 NOI/ICPC programming competition problems. The BenchPress row follows score sources that report OJBench (Pass@1), so the source-backed model-generation count is 232 rather than Pass@8 or dual-language variants."",""range"":[0,100],""sampling"":""Pass@1 for the BenchPress row; official paper also reports Pass@8 in separate settings"",""tools"":""code execution"",""version"":""OJBench Pass@1""}"
omnidocbench,"OmniDocBench (normalized edit distance, lower is better)",Vision,edit distance (lower=better),1651.0,https://huggingface.co/datasets/opendatalab/OmniDocBench,"{""higher_is_better"":false,""metric_type"":""normalized_edit_distance"",""multimodal_input"":true,""notes"":""Official OmniDocBench v1.6 contains 1,651 PDF pages. Count one model output per page for document parsing; HF parquet row count may differ slightly, but official README/page count is canonical. Stored scores use normalized edit distance on [0,1]; Kimi K3 reports 1-NED percentages, which are converted back to NED."",""range"":[0,1],""tools"":""none"",""version"":""OmniDocBench v1.6 full benchmark""}"
omnidocbench_1.5,OmniDocBench 1.5,Vision,normalized edit distance (lower=better),1355.0,https://github.com/opendatalab/OmniDocBench,"{""higher_is_better"":false,""metric_type"":""normalized_edit_distance"",""multimodal_input"":true,""notes"":""Average normalized edit distance; lower is better. The 1,355-page count is inferred from the official v1.6 update history."",""range"":[0,1],""tools"":""none"",""version"":""OmniDocBench v1.5""}"
omnimath,OmniMath,Math,,,https://arxiv.org/abs/2410.07985,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Olympiad-level math benchmark."",""range"":[0,100],""version"":""OmniMath""}"
osworld,OSWorld,Agentic,% success,369.0,https://os-world.github.io/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""tools=agentic (scaffold-defined). Preferred = official harness tools (bash/editor for SWE-bench, browser for OSWorld/BrowseComp, official APIs for tau-bench/MCPAtlas/Toolathlon, terminal for Terminal-Bench). Non-official scaffolds \u2192 matches_canonical=false."",""range"":[0,100],""tools"":""agentic"",""version"":""OSWorld (369)""}"
ovbench,OVBench,Video,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""OVBench""}"
ovobench,OVO-Bench,Streaming Video,aggregate online-video score (%),2814.0,https://arxiv.org/abs/2501.05510,"{""harness"":""official"",""higher_is_better"":true,""judge"":""task-specific rule/timing evaluation"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""2,814 meta-annotations over 644 videos."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""OVO-Bench""}"
paperbench,PaperBench,Coding,,,https://arxiv.org/abs/2507.20534,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Code dev from papers."",""range"":[0,100],""version"":""PaperBench Code-Dev""}"
phibench,PhiBench (Microsoft Internal),General,,,https://arxiv.org/abs/2412.08905,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Microsoft Phi team internal eval."",""range"":[0,100],""version"":""PhiBench 2.21 (Microsoft internal)""}"
phybench,Phybench,Physics,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""Phybench""}"
phyx_openended,PhyX (open-ended),Vision STEM,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""PhyX (open-ended)""}"
point_bench,Point-Bench,Vision Counting,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""Point-Bench""}"
popqa,PopQA,QA,,14267.0,https://huggingface.co/datasets/akariasai/PopQA,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official PopQA HuggingFace dataset contains 14,267 test rows. Count one factual QA generation per row; do not use rounded 14k marketing count."",""range"":[0,100],""tools"":""none"",""version"":""PopQA test set""}"
procbench,ProcBench,Reasoning,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""ProcBench""}"
realworldqa,RealWorldQA,Vision Perception,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""RealWorldQA""}"
refspatialbench,RefSpatialBench,Vision Spatial,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""RefSpatialBench""}"
repoqa,RepoQA,Coding,,500.0,https://arxiv.org/abs/2406.06025,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""RepoQA contains 500 code-search tasks from 50 repositories across 5 languages. Count task instances rather than repositories or candidate functions."",""range"":[0,100],""tools"":""none"",""version"":""RepoQA SNF, 32K context, threshold 0.8""}"
researchrubrics,ResearchRubrics,Deep Research,weighted rubric compliance score (%),101.0,https://huggingface.co/datasets/ScaleAI/researchrubrics/tree/85de3115053d1453ed612caacf4a405edc1ad756,"{""harness"":""official ResearchRubrics evaluation pipeline"",""higher_is_better"":true,""judge"":""binary rubric satisfaction with positive-weight average"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The pinned processed_data.jsonl contains 101 research tasks. Each report is scored against weighted binary rubrics."",""range"":[0,100],""sampling"":""one report per task"",""tools"":""research tools"",""version"":""ResearchRubrics official 101-task release""}"
ruler_128k,RULER 128K,Long Context,accuracy (%),6500.0,https://github.com/NVIDIA/RULER,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""13 tasks times 500 generated examples."",""range"":[0,100],""tools"":""none"",""version"":""RULER v1 13-task suite at 128K""}"
ruler_32k,RULER 32K,Long Context,accuracy (%),6500.0,https://github.com/NVIDIA/RULER,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""13 tasks times 500 generated examples."",""range"":[0,100],""tools"":""none"",""version"":""RULER v1 13-task suite at 32K""}"
safety,Safety (OLMES suite),Safety,,,https://arxiv.org/abs/2501.00656,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Allen AI internal safety eval suite."",""range"":[0,100],""version"":""OLMES safety suite""}"
scicode,SciCode,Coding,% correct,338.0,https://scicode-bench.github.io/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""SciCode contains 338 executable scientific-code subproblems. Count subproblems because each requires a code solution evaluated by tests."",""range"":[0,100],""tools"":""code execution"",""version"":""SciCode full subproblem benchmark""}"
screenspot_pro,ScreenSpot-Pro,Multimodal,,1581.0,https://github.com/likaixin2000/ScreenSpot-Pro-GUI-Grounding,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""ScreenSpot-Pro contains 1,581 GUI grounding targets: 604 icon targets and 977 text targets. Count one model grounding response per target."",""range"":[0,100],""tools"":""none"",""version"":""ScreenSpot-Pro full benchmark""}"
seal_0,Seal-0,Search Agent,%,,https://huggingface.co/moonshotai/Kimi-K2.5,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Kimi K2.5 model card."",""range"":[0,100],""tools"":""agentic"",""version"":""Seal-0""}"
sfe,SFE,Vision STEM,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""SFE""}"
simplebench,SimpleBench,Reasoning,% correct,1000.0,https://simple-bench.com/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""tools"":""none"",""version"":""SimpleBench (1000)""}"
simpleqa,SimpleQA,Knowledge,% correct,4326.0,https://openai.com/index/introducing-simpleqa/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""tools"":""none"",""version"":""SimpleQA (OpenAI 4326 questions)""}"
simpleqa_verified,SimpleQA-Verified,Knowledge,% correct (pass@1),,https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""tools"":""none"",""version"":""SimpleQA-Verified""}"
simplevqa,SimpleVQA,Vision,accuracy (%),2025.0,https://huggingface.co/datasets/m-a-p/SimpleVQA/tree/037cf89fb6f1212691756b66d1ecde6c2ce89e54,"{""harness"":""official SimpleVQA"",""higher_is_better"":true,""judge"":""LLM-as-judge"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""The pinned official simpleVQA_final_modified.json contains 2,025 items. Tool-assisted observations remain score-level non-default settings."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""SimpleVQA official 2,025-item test release""}"
smt_2025,SMT 2025,Math,% correct (pass@1),53.0,https://huggingface.co/datasets/MathArena/smt_2025,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""sampling"":""samples=4"",""tools"":""none"",""version"":""SMT 2025""}"
spreadsheetbench_verified,SpreadsheetBench Verified,Coding,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""code execution"",""version"":""SpreadsheetBench Verified""}"
superchem,Superchem (text-only),Chemistry,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""Superchem (text-only)""}"
supergpqa,SuperGPQA,Science,%,26529.0,https://huggingface.co/datasets/m-a-p/SuperGPQA,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official SuperGPQA dataset has 26,529 rows/questions spanning science disciplines. Count one answer per row."",""range"":[0,100],""tools"":""none"",""version"":""SuperGPQA full benchmark""}"
swe_bench_multilingual,SWE-bench Multilingual,Coding,% resolved (pass@1),300.0,https://huggingface.co/datasets/SWE-bench/SWE-bench_Multilingual/tree/846e647b9f33c0b51b739d005d13d85493c9af09,"{""harness"":""source-reported fixed coding scaffold"",""higher_is_better"":true,""judge"":""language-appropriate isolated test suites"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The pinned official dataset contains 300 issues across nine languages. The StepFun label SWE-MTLG is mapped to this released identity; scaffold remains score-level provenance."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""repository shell/editor agent"",""version"":""SWE-bench Multilingual official 300-instance test set""}"
swe_bench_multimodal,SWE-bench Multimodal,Coding,% resolved,,https://www.swebench.com/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""tools=agentic. No single standard public scaffold exists for SWE-bench Multimodal; harness choice is model-side (recorded in cell.reported_setting.harness). Any lab-published harness counts as canonical."",""range"":[0,100],""tools"":""agentic"",""version"":""SWE-bench Multimodal""}"
swe_bench_pro,SWE-bench Pro,Coding,% resolved (pass@1),731.0,https://huggingface.co/datasets/ScaleAI/SWE-bench_Pro/tree/7ab5114912baf22bb098818e604c02fe7ad2c11f,"{""harness"":""source-reported fixed coding scaffold"",""higher_is_better"":true,""judge"":""isolated repository test suites"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The pinned official public release contains 731 instances. Scaffold and effort remain score-level settings."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""repository shell/editor agent"",""version"":""SWE-bench Pro public 731-instance test set""}"
swe_bench_verified,SWE-bench Verified,Coding,% resolved (pass@1),500.0,https://huggingface.co/datasets/SWE-bench/SWE-bench_Verified/tree/78f471bf655a3137b2e8a75af1501690ec009ec3,"{""harness"":""source-reported fixed coding scaffold"",""higher_is_better"":true,""judge"":""isolated repository test suites"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The pinned official Verified release contains 500 instances. Scaffold, effort, Advisor mode, and internal reproductions remain score-level settings."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""repository shell/editor agent"",""version"":""SWE-bench Verified official 500-instance test set""}"
swe_evo,SWE-Evo,Agentic Coding,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""code execution"",""version"":""SWE-Evo""}"
swelancer,SWE-Lancer IC Diamond,Coding,%,198.0,https://github.com/openai/frontier-evals/tree/main/project/swelancer,"{""higher_is_better"":true,""judge"":""end-to-end tests"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Current official SWE-Lancer release contains 198 verified-offline IC SWE Diamond tasks; original paper reported 237 IC SWE Diamond tasks and the current README says 39 were dropped. Excludes SWE Manager Diamond."",""range"":[0,100],""sampling"":""pass@1, one attempt per task"",""tools"":""agentic"",""version"":""SWE-Lancer IC SWE Diamond, current verified offline release""}"
swelancer_freelance_dollars,SWE-Lancer IC SWE Diamond Freelance ($),Coding,dollars,198.0,https://github.com/openai/frontier-evals/tree/main/project/swelancer,"{""higher_is_better"":true,""judge"":""end-to-end tests"",""metric_type"":""dollars"",""multimodal_input"":false,""notes"":""Current official SWE-Lancer release contains 198 verified-offline IC SWE Diamond tasks; original paper reported 237 IC SWE Diamond tasks and the current README says 39 were dropped. Excludes SWE Manager Diamond."",""range"":[0,200000],""sampling"":""pass@1, one attempt per task"",""tools"":""agentic"",""version"":""SWE-Lancer IC SWE Diamond Freelance ($), current verified offline release""}"
tau1_bench_avg,"τ-bench (Yao 2024, averaged)",Tool use,,,https://cohere.com/research/papers/command-a-technical-report.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""\u03c4-bench (Yao 2024) averaged. Distinct from per-domain cells (tau_bench_retail/airline/telecom)."",""range"":[0,100],""version"":""\u03c4-bench averaged across retail+airline domains""}"
tau2_bench_airline,τ²-bench Airline,Agentic,% success,50.0,https://arxiv.org/abs/2506.07982,"{""higher_is_better"":true,""judge"":""state-based task success"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Paper Table 1 and current official split file both give 50 Airline tasks (30 train + 20 test). Dual-control text setting: LLM-controlled agent and simulated user; not comparable to original tau-bench."",""range"":[0,100],""sampling"":""pass^1 / one trial per task"",""tools"":""agentic"",""version"":""\u03c4\u00b2-bench (Sierra AI 2025) \u2014 airline domain, base split""}"
tau2_bench_avg,τ²-Bench (avg of retail/airline/telecom),Tool Use,macro task success (%),279.0,https://arxiv.org/abs/2506.07982,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Unweighted macro average over the three domain success rates; 50 + 115 + 114 underlying tasks."",""range"":[0,100],""tools"":""agentic"",""version"":""Tau2-bench macro average of airline, retail and telecom""}"
tau2_bench_retail,τ²-bench Retail,Agentic,% success,115.0,https://arxiv.org/abs/2506.07982,"{""higher_is_better"":true,""judge"":""state-based task success"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Paper Table 1 reports 115 Retail tasks. Current official repo base split has 114 after later task-fix releases; keep paper count for the tau2-bench 2025 row unless the row is redefined to current-release tau3 semantics."",""range"":[0,100],""sampling"":""pass^1 / one trial per task"",""tools"":""agentic"",""version"":""\u03c4\u00b2-bench (Sierra AI 2025) \u2014 retail domain, paper-defined task set""}"
tau2_bench_telecom,τ²-bench Telecom,Agentic,% success,114.0,https://arxiv.org/abs/2506.07982,"{""higher_is_better"":true,""judge"":""state-based task success"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Paper Table 1 and current official split file give 114 Telecom base tasks; the full generated Telecom pool has 2285 tasks and is excluded."",""range"":[0,100],""sampling"":""pass^1 / one trial per task"",""tools"":""agentic"",""version"":""\u03c4\u00b2-bench (Sierra AI 2025) \u2014 telecom domain, base split""}"
tau3_bench,τ³-Bench,Tool Use,%,1500.0,https://z.ai/blog/glm-5.1,"{""higher_is_better"":true,""judge"":""state-based task success"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Z.ai footnote says tau3-bench uses all domains with an extra user-simulator prompt, banking terminal_use retrieval, GPT-5.2-low user simulator, and 4 trials. Count = (airline 50 + retail 114 + telecom 114 + banking_knowledge 97) * 4 = 1500 scored simulations."",""range"":[0,100],""sampling"":""4 trials; count folded into num_problems"",""tools"":""agentic"",""version"":""\u03c4\u00b3-Bench all-domain text setting reported by Z.ai GLM-5.1 blog""}"
tau_bench_airline,tau-bench Airline,Agentic,% success,50.0,https://arxiv.org/abs/2406.12045,"{""higher_is_better"":true,""judge"":""state-based task success"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Original tau-bench Airline has 50 user-agent interaction tasks. Count task conversations; multi-turn/tool burden is represented by agentic cost factors rather than multiplying by every action step."",""range"":[0,100],""sampling"":""pass^1 / one trial per task"",""tools"":""agentic"",""version"":""tau-bench airline domain""}"
tau_bench_retail,Tau-Bench Retail,Agentic,% success,115.0,https://arxiv.org/abs/2406.12045,"{""higher_is_better"":true,""judge"":""state-based task success"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Original tau-bench Retail has 115 user-agent interaction tasks. Count task conversations; multi-turn/tool burden is represented by agentic cost factors rather than multiplying by every action step."",""range"":[0,100],""sampling"":""pass^1 / one trial per task"",""tools"":""agentic"",""version"":""tau-bench retail domain""}"
tau_bench_telecom,Tau-Bench Telecom,Agentic,% success,,https://arxiv.org/abs/2406.12045,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=agentic (scaffold-defined). Preferred = official harness tools (bash/editor for SWE-bench, browser for OSWorld/BrowseComp, official APIs for tau-bench/MCPAtlas/Toolathlon, terminal for Terminal-Bench). Non-official scaffolds \u2192 matches_canonical=false."",""range"":[0,100],""tools"":""agentic"",""version"":""tau-bench telecom""}"
tempcompass,TempCompass,Video,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""TempCompass""}"
terminal_bench,Terminal-Bench 2.0,Agentic,% tasks solved,445.0,https://arxiv.org/html/2601.11868v1,"{""harness"":""source-reported terminal scaffold"",""higher_is_better"":true,""judge"":""programmatic end-to-end task tests"",""logical_tasks"":89,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The immutable arXiv v1 paper defines 89 logical Terminal-Bench 2.0 tasks and evaluates every supported model-agent combination at least five times. num_problems therefore records 445 required generations. The paper's Table 2 token-count note refers to a 74-task execution subset and does not redefine the benchmark total. The StepFun mixed row maps Kimi K2.6, GPT-5.5, and Claude Opus 4.7 to this identity per the locked footnote."",""range"":[0,100],""required_trials_per_task"":5,""sampling"":""at least 5 trials per logical task"",""tools"":""terminal agent scaffold"",""version"":""Terminal-Bench 2.0 official 89-task release""}"
terminal_bench_1,Terminal-Bench 1.0,Agentic,% solved,,https://terminal-bench.com/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=agentic (scaffold-defined). Preferred = official harness tools (bash/editor for SWE-bench, browser for OSWorld/BrowseComp, official APIs for tau-bench/MCPAtlas/Toolathlon, terminal for Terminal-Bench). Non-official scaffolds \u2192 matches_canonical=false."",""range"":[0,100],""tools"":""agentic"",""version"":""Terminal-Bench 1.0""}"
terminal_bench_hard,Terminal-Bench Hard,Coding,%,,https://z.ai/blog/glm-4.7,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per GLM-4.7 blog."",""range"":[0,100],""tools"":""agentic"",""version"":""Terminal-Bench Hard""}"
tob_complex_workflows,ToB-ComplexWorkflows,Real-world,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""ToB-ComplexWorkflows""}"
tob_compositional_tasks,ToB-CompositionalTasks,Real-world,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""ToB-CompositionalTasks""}"
tob_information_extraction,ToB-InformationExtraction,Real-world,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""ToB-InformationExtraction""}"
tob_k12_education,ToB-K12Education,Real-world,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""ToB-K12Education""}"
tob_referenceqa,ToB-ReferenceQ&A,Real-world,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""ToB-ReferenceQ&A""}"
tob_text_classification,ToB-TextClassification,Real-world,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""ToB-TextClassification""}"
tomato,TOMATO,Video,multiple-choice temporal-reasoning accuracy (%),1484.0,https://arxiv.org/abs/2410.23266,"{""harness"":""official"",""higher_is_better"":true,""judge"":""exact multiple-choice grader"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""1,484 questions over 1,417 videos."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""TOMATO""}"
toolathlon,Toolathlon (Original),Agentic,% correct (pass@1),108.0,https://toolathlon.xyz/,"{""higher_is_better"":true,""judge"":""dedicated deterministic state evaluators"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Original 108-task benchmark across 32 applications and 604 tools. Distinct from Toolathlon-Verified introduced on 2026-06-30 with revised tasks/evaluators."",""range"":[0,100],""tools"":""agentic tool use"",""version"":""Original Toolathlon 108-task release before 2026-06-30""}"
treebench,TreeBench,Vision Spatial,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""TreeBench""}"
truthfulqa,TruthfulQA,Factuality,,817.0,https://github.com/sylinrl/TruthfulQA,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""TruthfulQA contains 817 questions designed to test imitative falsehoods. Count one text generation per question."",""range"":[0,100],""tools"":""none"",""version"":""TruthfulQA generation benchmark""}"
tvbench,TVBench,Video,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""TVBench""}"
usamo_2025,USAMO 2025,Math,% of 42 points,6.0,https://huggingface.co/datasets/MathArena/usamo_2025,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""tools"":""none"",""version"":""USAMO 2025""}"
usamo_2026,USAMO 2026,Math,% of 42 points,6.0,https://matharena.ai/usamo/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""tools=none preferred (pure-reasoning eval). If only with-tool scores (python/web/RAG) are available, accept and mark cell matches_canonical=false."",""range"":[0,100],""tools"":""none"",""version"":""USAMO 2026""}"
vending_bench_2,Vending-Bench 2,Agentic,,15000.0,https://andonlabs.com/evals/vending-bench-2,"{""higher_is_better"":true,""judge"":""year-end bank account balance"",""metric_type"":""dollars"",""multimodal_input"":false,""notes"":""Official Vending-Bench 2 reports leaderboard scores as the average across 5 full-year simulation runs. The page states that running a model for a full year results in 3,000-6,000 messages total, so this cost count uses the source-backed lower bound of actual model messages: 5 runs times 3,000 messages = 15,000. The true per-model count can be up to 30,000 messages; either way the benchmark is Tier 3."",""range"":[0,100000],""sampling"":""5 runs; lower-bound messages folded into num_problems"",""tools"":""agentic browser/email/order-management tools"",""version"":""Vending-Bench 2 (long-horizon planning)""}"
vibe_eval,Vibe-Eval,Multimodal,,269.0,https://github.com/reka-ai/reka-vibe-eval,"{""higher_is_better"":true,""judge"":""Reka Core evaluator scores each response on a 1-5 scale"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Official paper and HF dataset report 269 visual-understanding prompts, including 100 hard prompts. Count model generations as one response per example_id; evaluator calls are scoring overhead."",""range"":[0,100],""sampling"":""single response per prompt"",""version"":""Vibe-Eval v1 overall""}"
video_mme,Video-MME,Multimodal,% multiple-choice accuracy,2700.0,https://raw.githubusercontent.com/MME-Benchmarks/Video-MME/06c2315b892f88578f81d73205d07cf576f292b9/README.md,"{""higher_is_better"":true,""judge"":""answer-key multiple-choice accuracy"",""metric"":""multiple-choice QA accuracy"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Official immutable repository defines 900 videos and 2,700 human-annotated question-answer pairs."",""range"":[0,100],""sampling"":""one response per question"",""tools"":""none"",""version"":""Video-MME overall, 900 videos / 2,700 QA pairs""}"
video_mmmu,Video-MMMU,Video/Multimodal,%,900.0,https://videommmu.github.io/,"{""higher_is_better"":true,""metric"":""accuracy over human-annotated video QA questions"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Official paper/project report 300 expert-level videos and 900 human-annotated questions across Perception, Comprehension, and Adaptation. Count one model generation per question for the overall percent score; do not count the model-report source page as the benchmark definition."",""range"":[0,100],""sampling"":""single response per question"",""tools"":""none"",""version"":""Video-MMMU overall""}"
videoeval_pro,VideoEval-Pro,Video,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""VideoEval-Pro""}"
videoholmes,VideoHolmes,Video,% multiple-choice accuracy,1837.0,https://raw.githubusercontent.com/TencentARC/Video-Holmes/52ef8da286ccad03036a65e7b67c160bc3a24fb9/README.md,"{""higher_is_better"":true,""judge"":""answer-key multiple-choice accuracy"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Official immutable repository defines 1,837 questions from 270 suspense short films across seven tasks."",""range"":[0,100],""sampling"":""one response per question"",""tools"":""none"",""version"":""Video-Holmes full evaluation set""}"
videoreasonbench,VideoReasonBench,Video,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""VideoReasonBench""}"
videosimpleqa,VideoSimpleQA,Video,source-reported score (%),1504.0,https://huggingface.co/datasets/VideoSimpleQA/VideoSimpleQA,"{""harness"":""official"",""higher_is_better"":true,""judge"":""official configurable LLM grader"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""1,504 QA pairs over 1,079 videos. The Seed source does not identify whether the displayed score is the official accuracy or F1 view."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""VideoSimpleQA""}"
visfactor,VisFactor,Vision Perception,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""VisFactor""}"
vispeak,ViSpeak,Video,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""ViSpeak""}"
visulogic,VisuLogic,Vision Puzzles,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""VisuLogic""}"
vitabench,VitaBench,Tool Use,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""tool calls"",""version"":""VitaBench""}"
viverbench,ViVerBench,Vision VQA,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""ViVerBench""}"
vlms_are_biased,VLMsAreBiased,Vision Perception,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""VLMsAreBiased""}"
vlms_are_blind,VLMsAreBlind,Vision Perception,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""VLMsAreBlind""}"
vpct,VPCT,Vision Puzzles,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""VPCT""}"
widesearch,WideSearch (item-F1),Search Agent,%,200.0,https://huggingface.co/datasets/ByteDance-Seed/WideSearch,"{""higher_is_better"":true,""judge"":""evaluation pipeline mixes exact/URL/numeric matching with LLM-assisted cell judgment"",""metric"":""item-F1 over required table fields"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official paper and dataset report 200 broad information-seeking tasks, split 100 English and 100 Chinese. Count one model generation per task; item-F1 expands the scoring units, not the number of model generations."",""range"":[0,100],""sampling"":""single final response per task"",""tools"":""search/browser access required"",""version"":""WideSearch overall (item-F1)""}"
wildbench,WildBench,Chat,Raw Score,,https://github.com/InternLM/InternLM,"{""higher_is_better"":true,""metric_type"":""raw"",""multimodal_input"":false,""notes"":""Per InternLM3 GitHub README. WildBench raw score."",""range"":[null,null],""tools"":""none"",""version"":""WildBench (Raw Score)""}"
world_travel_text,WorldTravel (TEXT),Real-world,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""WorldTravel (TEXT)""}"
world_travel_vlm,WorldTravel (VLM),Real-world,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""WorldTravel (VLM)""}"
worldvqa,WorldVQA,Multimodal Knowledge,accuracy (%),3000.0,https://huggingface.co/datasets/moonshotai/WorldVQA/tree/29e1d54b27ffb34cdffb4cdc95d29afcf101f1f7,"{""harness"":""official"",""higher_is_better"":true,""judge"":""official default gpt-oss-120b judge"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Released first-eight-category leaderboard excluding People."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""WorldVQA""}"
xbench_deepsearch,xbench-DeepSearch,Search Agent,%,100.0,https://huggingface.co/datasets/xbench/DeepSearch,"{""higher_is_better"":true,""metric"":""accuracy"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official xbench DeepSearch HF dataset reports 100 encrypted rows/tasks and describes a search/information-retrieval evaluation. Count one model generation per task; do not use the MiniMax M2 model card as the benchmark definition. Later DeepSearch-2510 is a separate variant and also has 100 rows."",""range"":[0,100],""sampling"":""single answer per task"",""tools"":""search/retrieval environment required"",""version"":""xbench-DeepSearch original release""}"
xlrs_macro,XLRS-Bench (macro),Vision STEM,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""XLRS-Bench (macro)""}"
xpert_bench,XPertBench,Economic,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""XPertBench""}"
zerobench_main,ZeroBench (main),Vision Puzzles,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""ZeroBench (main)""}"
zerobench_sub,ZeroBench (sub),Vision Puzzles,%,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Per Doubao Seed 2.0 Pro model card."",""range"":[0,100],""tools"":""none"",""version"":""ZeroBench (sub)""}"
zerobench_tools,ZeroBench main (with tools),Multimodal Reasoning,accuracy (%),100.0,https://arxiv.org/abs/2502.09696,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Main 100-question set with tool access."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""search/code/web tools"",""version"":""ZeroBench main (with tools)""}"
kimi_code_bench_v2,Kimi Code Bench v2,Coding,score (%),,https://huggingface.co/moonshotai/Kimi-K2.7-Code,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Moonshot in-house coding-agent benchmark spanning 10+ languages and production software-engineering tasks. The source does not report the task count."",""range"":[0,100],""version"":""Kimi Code Bench v2""}"
program_bench,ProgramBench,Coding,macro-average behavioral tests passed (%),200.0,https://programbench.com/,"{""harness"":""official ProgramBench sandbox"",""higher_is_better"":true,""judge"":""248,000+ fuzz-generated behavioral tests"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Canonical score is the macro-average behavioral-tests-passed rate. Full task resolution is a distinct future metric and is not mixed into this id."",""range"":[0,100],""sampling"":""unknown"",""tools"":""compiled executable and documentation; no source, decompilation, or internet"",""version"":""ProgramBench 200-task suite""}"
mls_bench_lite,MLS-Bench-Lite,Coding,score (0-100),30.0,https://mls-bench.com/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official 30-task subset of MLS-Bench. Agents receive five hours to develop and submit scalable ML methods."",""range"":[0,100],""version"":""Official MLS-Bench-Lite 30-task subset""}"
kimi_claw_24_7,Kimi Claw 24/7 Bench,Agentic,% average pass rate,17.0,https://huggingface.co/moonshotai/Kimi-K2.7-Code,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""17 persistent multi-day professional scenarios covering 610 evaluation points in the OpenClaw harness. Final score is the average pass rate over evaluation points and three runs."",""range"":[0,100],""version"":""Kimi Claw 24/7 Bench""}"
mcpmark_verified,MCPMark-Verified,Agentic,% success,,https://huggingface.co/moonshotai/Kimi-K2.7-Code,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Human-verified edition across Notion, GitHub, Filesystem, Postgres, and Playwright. Official configuration uses a 100-step tool-call budget, 32K max tokens per step, and averages three runs. Task count was not reported."",""range"":[0,100],""version"":""MCPMark-Verified human-verified edition""}"
aa_briefcase_elo,AA-Briefcase (Elo),Agentic,Elo rating,91.0,https://artificialanalysis.ai/evaluations/aa-briefcase,"{""higher_is_better"":true,""metric_type"":""elo_rating"",""multimodal_input"":true,""notes"":""Composite Artificial Analysis rating over rubric correctness, analytical quality, and presentation quality; live snapshot scores can drift."",""range"":null,""version"":""AA-Briefcase live 91-task series""}"
agents_last_exam,Agents' Last Exam,Agentic,% tasks passed,,https://agents-last-exam.org/docs/ale/index.html,"{""higher_is_better"":true,""metric_type"":""pass_rate_pct"",""multimodal_input"":true,""notes"":""Professional computer-use tasks with hidden-reference grading. The public corpus is growing, so the scored item count is not stable."",""range"":[0,100],""version"":""Living public benchmark; no fixed scored version""}"
automation_bench,AutomationBench,Agentic,% tasks passed,600.0,https://github.com/zapier/AutomationBench,"{""higher_is_better"":true,""metric_type"":""strict_pass_rate_pct"",""multimodal_input"":false,""notes"":""A task passes only when every final-state assertion passes. Distinct from AutomationBench-AA."",""range"":[0,100],""version"":""Public 600-task scored set""}"
corpfin_v2,CorpFin v2,Finance,% accuracy,858.0,https://www.vals.ai/benchmarks/corp_fin_v2,"{""higher_is_better"":true,""metric_type"":""accuracy_pct"",""multimodal_input"":false,""notes"":""858 questions from 43 credit agreements, evaluated under documented context variants."",""range"":[0,100],""version"":""CorpFin v2 held-out test set""}"
deep_swe_v1_1,DeepSWE v1.1,Agentic Coding,% resolved (pass@1),113.0,https://github.com/datacurve-ai/deep-swe,"{""harness"":""Pier newer than 0.3.0 with a separate pristine verifier environment"",""higher_is_better"":true,""judge"":""program-based functional and regression verifiers"",""metric_type"":""pass_at_1_pct"",""multimodal_input"":false,""notes"":""Official 113-task v1.1 corpus across five languages. The official leaderboard uses Pier and mini-swe-agent; Meta's chart uses selected agent products and is not leaderboard-harness-identical."",""range"":[0,100],""sampling"":""pass@1; repeat count is score-level provenance"",""tools"":""agentic repository shell/editor"",""version"":""DeepSWE v1.1""}"
finance_agent_v2,Finance Agent v2,Finance,% weighted partial credit,927.0,https://www.vals.ai/benchmarks/fabv2,"{""higher_is_better"":true,""judge"":""three-model jury: GPT-5.4, Gemini 3.1 Pro, Claude Sonnet 4.6"",""metric_type"":""weighted_partial_credit_pct"",""multimodal_input"":false,""notes"":""Finance Agent v2 has 927 questions: 27 public, 450 private validation, 450 held-out test. Published scores use the hidden 450-task test and dealbreaker-gated severity-weighted partial credit."",""range"":[0,100],""sampling"":""3 runs/model on 450 held-out scored tasks"",""tools"":""Vals six-tool finance-agent harness"",""version"":""Finance Agent v2""}"
frontier_swe,FrontierSWE,Agentic Coding,dominance (%),17.0,https://www.frontierswe.com/,"{""harness"":""public FrontierSWE harness"",""higher_is_better"":true,""judge"":""continuous partial-credit task scoring and dominance"",""metric_type"":""dominance_pct"",""multimodal_input"":false,""notes"":""Dominance is win probability against a random opponent over continuous task scores; Grok 4.5 uses Grok CLI."",""range"":[0,100],""sampling"":""mean@5"",""tools"":""agentic code execution"",""version"":""FrontierSWE initial 17-task release; 20-hour budget""}"
harvey_lab_aa,Harvey LAB-AA,Agentic,% rubric criteria passed,120.0,https://artificialanalysis.ai/evaluations/harvey-lab-aa,"{""higher_is_better"":true,""metric_type"":""criterion_pass_rate_pct"",""multimodal_input"":true,""notes"":""Criterion pass rate over 120 private legal tasks across 24 practice areas."",""range"":[0,100],""version"":""Artificial Analysis 120-task LAB-AA implementation""}"
job_bench,JobBench,Agentic,% weighted rubric score,65.0,https://github.com/Job-Bench/job-bench-eval,"{""higher_is_better"":true,""metric_type"":""weighted_rubric_score_pct"",""multimodal_input"":true,""notes"":""Professional-work deliverables scored by weighted rubrics; excludes the easy smoke-test split."",""range"":[0,100],""version"":""Main 65-task leaderboard split""}"
legal_research_bench,Legal Research Bench,Agentic,% all-pass accuracy,,https://www.vals.ai/benchmarks/legal_research,"{""higher_is_better"":true,""metric_type"":""all_pass_accuracy_pct"",""multimodal_input"":false,""notes"":""A question passes only when every required rubric item passes; exact hidden count is not public."",""range"":[0,100],""version"":""Current Vals held-out suite""}"
osworld_2_0,OSWorld 2.0,Agentic,% weighted checkpoint score,108.0,https://osworld-v2.xlang.ai/,"{""higher_is_better"":true,""metric_type"":""weighted_checkpoint_score_pct"",""multimodal_input"":true,""notes"":""Separate 108-workflow benchmark with weighted checkpoint partial scoring at the standard 500-step budget."",""range"":[0,100],""version"":""OSWorld 2.0""}"
osworld_verified,OSWorld-Verified,Agentic,% task success,369.0,https://xlang.ai/blog/osworld-verified,"{""higher_is_better"":true,""metric_type"":""task_success_rate_pct"",""multimodal_input"":true,""notes"":""Repaired OSWorld 1.x lineage, distinct from OSWorld 2.0. Some runs exclude eight Google Drive tasks; score-level notes must disclose exclusions when known."",""range"":[0,100],""version"":""OSWorld-Verified revision announced 2025-07-28""}"
perception_bench,PerceptionBench,Vision Perception,% accuracy,3000.0,https://github.com/MoonshotAI/PerceptionBench,"{""higher_is_better"":true,""metric_type"":""accuracy_pct"",""multimodal_input"":true,""notes"":""Open-ended visual perception questions across ten atomic capabilities; binary judge verdict per response."",""range"":[0,100],""version"":""Initial 3,000-question release""}"
posttrain_bench,PostTrainBench,Agentic Research,weighted average objective score (%),28.0,https://github.com/aisa-group/PostTrainBench,"{""harness"":""one H100 for 10 hours"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Seven objectives across four base models; count actual objective runs as 28."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""post-training pipeline modification"",""version"":""PostTrainBench""}"
saas_bench,SaaS-Bench,Agentic,% checkpoint score,106.0,https://github.com/UniPat-AI/SaaS-Bench,"{""higher_is_better"":true,""metric_type"":""checkpoint_score_pct"",""multimodal_input"":true,""notes"":""Workflow checkpoint score across 23 self-hosted SaaS applications; distinct from strict resolved-task rate."",""range"":[0,100],""version"":""Initial 106-task release""}"
spreadsheetbench_2,SpreadsheetBench 2,Office,% modification accuracy,321.0,https://github.com/RUCKBReasoning/SpreadsheetBench-2,"{""higher_is_better"":true,""metric_type"":""modification_accuracy_pct"",""multimodal_input"":true,""notes"":""Spreadsheet modification benchmark covering debugging, financial models, templates, and visualization."",""range"":[0,100],""version"":""SpreadsheetBench 2""}"
swe_marathon_h20_2026_07_09,SWE-Marathon H20 Snapshot (2026-07-09),Agentic Coding,% resolved (pass@1),20.0,https://github.com/abundant-ai/swe-marathon,"{""higher_is_better"":true,""metric_type"":""pass_at_1_pct"",""multimodal_input"":false,""notes"":""Source-specific Moonshot H20 calibration before final v1.1; kept separate from canonical SWE-Marathon releases."",""range"":[0,100],""version"":""H20-calibrated pre-final-v1.1 branch dated 2026-07-09""}"
tau3_banking,τ³-Banking,Tool Use,% passed (pass@1),97.0,https://github.com/sierra-research/tau2-bench,"{""higher_is_better"":true,""metric_type"":""pass_at_1_pct"",""multimodal_input"":false,""notes"":""Knowledge-retrieval and transactional banking customer-service scenarios. K3 source does not pin the corrected release."",""range"":[0,100],""version"":""\u03c4\u00b3-bench banking_knowledge""}"
terminal_bench_2_1,Terminal-Bench 2.1,Agentic Coding,% resolved (pass@1),89.0,https://www.tbench.ai/news/terminal-bench-2-1,"{""harness"":""Harbor with submitted agent/model/sandbox provenance"",""higher_is_better"":true,""judge"":""task executable verifier"",""metric_type"":""pass_at_1_pct"",""multimodal_input"":false,""notes"":""All 89 tasks. The official release page says 28 tasks changed from 2.0; the pinned official repository README says 26. The task count agrees and the changed-task-count conflict remains explicit."",""range"":[0,100],""sampling"":""pass@1; leaderboard submission requires at least five trials per task"",""tools"":""terminal agent in the official container task"",""version"":""Terminal-Bench 2.1""}"
toolathlon_verified,Toolathlon-Verified,Tool Use,mean pass@1 (%),108.0,https://toolathlon.xyz/docs/blog/toolathlon-verified,"{""higher_is_better"":true,""metric_type"":""mean_pass_at_1_pct"",""multimodal_input"":true,""notes"":""Official repaired release, separate from original Toolathlon; mean pass@1 across three runs."",""range"":[0,100],""version"":""Toolathlon-Verified released 2026-06-30""}"
coding_experience,Coding Experience,Coding,score (%),,https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Practical coding-agent experience in real development workflows; provider-aligned Claude Code, Kimi Code, or Codex harness."",""range"":[0,100],""version"":""Kimi internal Coding Experience; version unspecified""}"
clawbench_2_0,24/7 ClawBench 2.0,Agentic,score (%),,https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Always-on, multi-day assistant tasks with concurrent events and interruptions; OpenClaw harness."",""range"":[0,100],""version"":""24/7 ClawBench 2.0""}"
mira_bench,MIRA Bench,Agentic,score (%),,https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Multi-agent enterprise collaboration, delegation, and routing; MIRA harness."",""range"":[0,100],""version"":""Kimi internal MIRA Bench; version unspecified""}"
kaet,KAET,Agentic,score (%),,https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Long-horizon autonomous execution simulating user requests and enterprise operations; Kimi Code harness."",""range"":[0,100],""version"":""Kimi Autonomous Execution Tasks; version unspecified""}"
clif_bench,CLIF Bench,Agentic,score (%),,https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""In-context learning with instructions interleaving multiple complex skills; Kimi Code harness."",""range"":[0,100],""version"":""Context Learning and Instruction Following Bench; version unspecified""}"
agentic_vision_bench,Agentic Vision Bench,Vision Agent,score (%),,https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Tests whether agents notice and use key visual facts during execution; Kimi Code harness."",""range"":[0,100],""version"":""Kimi internal Agentic Vision Bench; version unspecified""}"
swarm_bench,SwarmBench,Agentic,score (%),,https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Agent-swarm orchestration through coordinated decomposition and parallel execution; Kimi Agent harness."",""range"":[0,100],""version"":""Kimi internal SwarmBench; version unspecified""}"
online_experience,Online Experience,Agentic,score (%),,https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Real online-agent usage and commonly requested deliverable file types; Kimi Agent harness."",""range"":[0,100],""version"":""Kimi internal Online Experience; version unspecified""}"
finance_bench,Finance Bench,Finance,score (%),,https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Kimi internal realistic financial work from source materials to reviewable deliverables; distinct from public FinanceBench."",""range"":[0,100],""version"":""Kimi internal Finance Bench; version unspecified""}"
kwv_bench,KWVBench,Vision Agent,score (%),,https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Atomic visual capabilities distilled from real knowledge-work scenarios."",""range"":[0,100],""version"":""Kimi internal Knowledge Work Vision Bench; version unspecified""}"
deck_bench,DECKBench,Office,score (%),,https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Presentation-deck generation from real-usage task descriptions."",""range"":[0,100],""version"":""Kimi internal DECKBench; version unspecified""}"
agent_behavior_bench,Agent Behavior Bench,Agentic,score (%),,https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Process quality, tool use, efficiency, and discipline alongside task completion; Kimi Work harness."",""range"":[0,100],""version"":""Kimi internal Agent Behavior Bench; version unspecified""}"
faithfulness,Faithfulness,Factuality,1 - hallucination rate (%),,https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Responses are fact-checked; the reported higher-is-better metric is one minus hallucination rate."",""range"":[0,100],""version"":""Kimi internal Faithfulness; version unspecified""}"
chat_all_in_one_bench,Chat All-in-One Bench,Chat,score (%),,https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Conversational experience across every stage of product usage; Kimi Work harness."",""range"":[0,100],""version"":""Kimi internal Chat All-in-One Bench; version unspecified""}"
aa_intelligence_index_v4_1,Artificial Analysis Intelligence Index v4.1,Composite,composite score,,https://artificialanalysis.ai/methodology/intelligence-benchmarking,"{""higher_is_better"":true,""metric_type"":""weighted_composite_score"",""multimodal_input"":false,""notes"":""Version-pinned English text-only composite. Stored scores are the K3 report's 2026-07-23 live leaderboard snapshot."",""range"":[0,100],""version"":""Artificial Analysis Intelligence Index v4.1""}"
vals_index,Vals Index,Composite,weighted accuracy (%),,https://www.vals.ai/benchmarks/vals_index,"{""higher_is_better"":true,""metric_type"":""weighted_accuracy_pct"",""multimodal_input"":false,""notes"":""GDP-weighted professional benchmark composite; methodology can change over time."",""range"":[0,100],""version"":""Live Vals Index snapshot""}"
webdev_arena_elo,WebDevArena Elo,Human Preference,Arena score,,https://arena.ai/leaderboard/code/webdev,"{""higher_is_better"":true,""metric_type"":""arena_score"",""multimodal_input"":false,""notes"":""Blind pairwise preference over generated web applications; scores drift as votes accumulate."",""range"":null,""version"":""Live leaderboard snapshot""}"
text_arena_elo,Arena Text Elo,Human Preference,Elo rating,,https://arena.ai/leaderboard/text,"{""higher_is_better"":true,""judge"":""blind human pairwise preference votes"",""metric_type"":""elo"",""multimodal_input"":false,""notes"":""Dynamic Elo snapshot; score observations must record snapshot date because ratings drift."",""range"":null,""tools"":""none"",""version"":""Live Arena Text leaderboard snapshot""}"
agent_arena,AgentArena,Agentic,net improvement,,https://arena.ai/blog/agent-arena-methodology,"{""higher_is_better"":true,""metric_type"":""mean_treatment_effect"",""multimodal_input"":false,""notes"":""Causal treatment-effect estimate over real Agent Mode sessions; no fixed task count or invariant baseline."",""range"":null,""version"":""Live leaderboard snapshot""}"
moonshot_exploit_development_suite,Moonshot Exploit Development Suite,Cyber,% tasks solved,36.0,https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf,"{""higher_is_better"":true,""metric_type"":""task_solve_rate_pct"",""multimodal_input"":false,""notes"":""16 user-space CVE exploitation tasks and 20 historical Linux-kernel CVE tasks in QEMU; human-verified solvable."",""range"":[0,100],""version"":""Kimi K3 report 36-task in-house exploit suite""}"
exploitbench,ExploitBench,Cyber,% capability success,41.0,https://arxiv.org/abs/2605.14153,"{""higher_is_better"":true,""metric_type"":""capability_ladder_success_pct"",""multimodal_input"":false,""notes"":""Public 41-vulnerability exploitation benchmark with deterministically verified capability ladders."",""range"":[0,100],""version"":""ExploitBench V8 41-task corpus (2026)""}"
dsbench_fullstack,DSBench-FullStack,Agentic Coding,score (%),,https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""DeepSeek internal full-stack development test set. The official model card does not report its task count."",""range"":[0,100],""version"":""DeepSeek internal DSBench-FullStack; version unspecified""}"
dsbench_hard,DSBench-Hard,Agentic Coding,score (%),,https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""DeepSeek internal test set of difficult coding-agent problems. The official model card does not report its task count."",""range"":[0,100],""version"":""DeepSeek internal DSBench-Hard; version unspecified""}"
hle_tools_text,HLE Text (w/ tools),Reasoning & Knowledge,accuracy (%),2158.0,https://z.ai/blog/glm-5.2,"{""higher_is_better"":true,""judge"":""answer-key scoring for closed-ended answers"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""GLM-5.2 explicitly marks unstarred HLE-with-tools values as text-only and starred values as the full text+image set. Text-only and full-set scores are distinct benchmark identities."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""search, code execution, and web browsing"",""version"":""HLE finalized text-only subset (2,158 questions) with tools""}"
swe_marathon_v1_0,SWE-Marathon v1.0,Agentic Coding,% resolved (pass@1),20.0,https://github.com/abundant-ai/swe-marathon/releases/tag/v1.0,"{""higher_is_better"":true,""metric_type"":""pass_at_1_pct"",""multimodal_input"":false,""notes"":""Official v1.0 evaluates 20 ultra-long-horizon tasks with five trials per agent-model-task pair. It predates and is distinct from the July v1.1/H20-calibrated snapshot already stored in BP."",""range"":[0,100],""sampling"":""trials=5; reported as pass@1"",""tools"":""agentic code execution"",""version"":""SWE-Marathon v1.0 released 2026-06-10""}"
hy_backend_2_0,Hy-Backend 2.0 (Internal),Agentic Coding,reported score (%),,https://huggingface.co/tencent/Hy3,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Tencent internal backend benchmark; public item count and metric definition are not published."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""Claude Code; GPT-5.5 uses CodeX"",""version"":""Hy-Backend 2.0 (Internal)""}"
hy_swe_max,Hy-SWE Max (Internal),Agentic Coding,reported score (%),,https://huggingface.co/tencent/Hy3,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Tencent internal software-engineering benchmark; public item count and metric definition are not published."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""Claude Code; GPT-5.5 uses CodeX"",""version"":""Hy-SWE Max (Internal)""}"
hy_company_bench,Hy-CompanyBench (Internal),Agentic Coding,reported score (%),,https://huggingface.co/tencent/Hy3,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Tencent internal company-task benchmark; public item count and metric definition are not published."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""Claude Code; GPT-5.5 uses CodeX"",""version"":""Hy-CompanyBench (Internal)""}"
wildclaw_bench_35_text,"WildClawBench (35, text-only)",Agentic,reported score (%),35.0,https://huggingface.co/tencent/Hy3,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official Hy3 footnote defines the text-only 35-query subset."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""OpenClaw harness"",""version"":""WildClawBench (35, text-only)""}"
skills_bench_text_79,"SkillsBench (79, text-only)",Agentic,reported score (%),79.0,https://huggingface.co/tencent/Hy3,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Self-contained 79-task subset; multimodal tasks excluded."",""range"":[0,100],""sampling"":""average over 3 runs"",""tools"":""Claude Code"",""version"":""SkillsBench (79, text-only)""}"
e_bench_internal,e-bench (Internal),Agentic,reported score (%),,https://huggingface.co/tencent/Hy3,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Tencent internal working-agent benchmark; public item count and protocol are not published."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""benchmark-specific agent tools"",""version"":""e-bench (Internal)""}"
hy_finmodel_bench,Hy-FinModelBench (Internal),Agentic,reported score (%),,https://huggingface.co/tencent/Hy3,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Tencent internal financial-modeling benchmark; public item count and protocol are not published."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""benchmark-specific agent tools"",""version"":""Hy-FinModelBench (Internal)""}"
prod_bench_internal,"ProdBench (Internal, pass^3)",Agentic,reported score (%),,https://huggingface.co/tencent/Hy3,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Tencent internal productivity benchmark evaluated with OpenClaw; public item count is not published."",""range"":[0,100],""sampling"":""pass^3"",""tools"":""OpenClaw harness"",""version"":""ProdBench (Internal, pass^3)""}"
hy_skillsworld,Hy-SkillsWorld (Internal),Agentic,reported score (%),,https://huggingface.co/tencent/Hy3,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Tencent internal skills benchmark; public item count and protocol are not published."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""benchmark-specific agent tools"",""version"":""Hy-SkillsWorld (Internal)""}"
hy_euler_pro,"Hy-Euler Pro (Internal, tools)",Reasoning & Knowledge,reported score (%),,https://huggingface.co/tencent/Hy3,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Tencent internal STEM-agent benchmark; public item count and metric definition are not published."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""benchmark-specific tools"",""version"":""Hy-Euler Pro (Internal, tools)""}"
horizon_math_pass12,HorizonMath (pass@12),Reasoning & Knowledge,reported score (%),113.0,https://github.com/ewang26/HorizonMath,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official HorizonMath repository defines 113 automatically verified research problems across eight domains."",""range"":[0,100],""sampling"":""pass@12"",""tools"":""none"",""version"":""HorizonMath (pass@12)""}"
hy_math_internal,Hy-Math (Internal),Reasoning & Knowledge,reported score (%),,https://huggingface.co/tencent/Hy3,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Tencent internal mathematics benchmark; public item count and protocol are not published."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""Hy-Math (Internal)""}"
cmt_benchmark,CMT-Benchmark,Reasoning & Knowledge,reported score (%),50.0,https://github.com/JamesRoggeveen/cmt_benchmark_data,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official CMT-Benchmark paper/repository defines 50 expert-authored condensed-matter problems."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""CMT-Benchmark""}"
cl_bench_life,CL-Bench Life,Long Context,reported score (%),405.0,https://huggingface.co/datasets/tencent/CL-bench-Life,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official Tencent dataset defines 405 context-task pairs and 5,348 evaluation rubrics."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""CL-Bench Life""}"
forte_avg3,FORTE (Avg@3),Agents,Avg@3 (%),180.0,https://github.com/AGI-Eval-Official/FORTE,"{""harness"":""OpenClaw in Docker"",""higher_is_better"":true,""judge"":""LLM-as-judge over expert rubrics; all-or-nothing per run"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Official README declares 180 full tasks across 15 professions. The official leaderboard JSON values appear arithmetically compatible with a 183-task denominator, an unresolved publisher inconsistency. LongCat used 45-minute task timeouts; public demo schemas use 40 minutes."",""range"":[0,100],""sampling"":""trials=3"",""tools"":""office-computing environment"",""version"":""FORTE June 2026""}"
rwsearch,RWSearch,Agents,publisher Score (0-100),200.0,https://github.com/AGI-Eval-Official/RW-Search,"{""context_management"":""none"",""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official repository defines 200 Chinese real-world search questions with unique objective answers. The publisher labels the metric only as Score and does not disclose the full matcher or aggregation."",""range"":[0,100],""tools"":""Search and Browse"",""version"":""RWSearch 2026""}"
advancedif,AdvancedIF,Instruction Following,overall pass rate (%),1645.0,https://arxiv.org/abs/2511.10507,"{""higher_is_better"":true,""judge"":""LLM-as-judge; public evaluator default o3-mini-2025-01-31"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""1,645 expert-rubric prompts across system steerability, carried context, and complex instruction following. Canonical score is the percentage of samples where all rubrics pass."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""AdvancedIF public test split""}"
xlrs_bench_micro,XLRS-Bench (micro),Vision STEM,micro-average (%),45942.0,https://arxiv.org/abs/2503.23771,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Full ultra-high-resolution remote-sensing benchmark with 45,942 annotations across 16 tasks. The official full-benchmark Avg. is a micro average; XLRS-Bench-lite uses a macro average."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""XLRS-Bench full""}"
microvqa,MicroVQA,Vision STEM,mean accuracy (%),1042.0,https://arxiv.org/abs/2503.13399,"{""higher_is_better"":true,""judge"":""rule-based multiple-choice accuracy"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""1,042 expert-curated microscopy multiple-choice questions covering perception, hypothesis generation, and experiment proposal."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""MicroVQA v0.0.1 test split""}"
sgi_bench,SGI-Bench,Scientific Agents,SGI-Score,,https://arxiv.org/abs/2512.16969,"{""harness"":""official SGI-Bench agentic evaluation"",""higher_is_better"":true,""judge"":""task-specific official metrics aggregated into SGI-Score"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Scientist-aligned full inquiry-cycle suite spanning 10 disciplines and more than 1,000 gated expert-curated samples. The publisher does not state one exact total item count."",""range"":[0,100],""tools"":""web search, PDF parser, Python interpreter, file reader"",""version"":""SGI-Bench full gated suite""}"
researchclawbench,ResearchClawBench,Scientific Agents,weighted rubric score (%),40.0,https://arxiv.org/abs/2606.07591,"{""harness"":""ResearchHarness"",""higher_is_better"":true,""judge"":""weighted task-specific rubrics"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""40 real-science research tasks across 10 domains."",""range"":[0,100],""tools"":""scientific research environment"",""version"":""ResearchClawBench core""}"
gdpval_normalized_elo,GDPVal (normalized Elo),Economic,normalized Elo (0-100),220.0,https://huggingface.co/datasets/openai/gdpval,"{""higher_is_better"":true,""judge"":""Gemini 3.1 Pro rubric judge"",""metric_type"":""index"",""multimodal_input"":false,""notes"":""Distinct from raw GDPVal Artificial Analysis Elo. NVIDIA reports normalized=(Elo-500)/2000 on a 0-100 display scale."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""office workflow, web search, sandboxed code"",""version"":""GDPVal public 220-task set; normalized Elo on 0-100 scale""}"
profbench,ProfBench (Search),Search Agent,search score,40.0,https://github.com/NVlabs/ProfBench,"{""higher_is_better"":true,""judge"":""rubric-based criterion grading"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Professional-domain rubric tasks with search and browsing."",""range"":[0,100],""sampling"":""16-run average"",""tools"":""web search and browsing"",""version"":""ProfBench full 40-task release""}"
pinchbench,PinchBench,Agentic,% passed,53.0,https://github.com/pinchbench/skill,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Open public subset evaluated in the OpenClaw environment."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""OpenClaw coding environment"",""version"":""PinchBench public 53-task release""}"
tau3_airline,tau3-bench Airline,Tool Use,% passed (8-trial average),400.0,https://github.com/sierra-research/tau2-bench,"{""higher_is_better"":true,""judge"":""state-based task success"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""50 tasks x 8 trials; extra simulator prompt and GPT-5.2-low simulator."",""range"":[0,100],""sampling"":""50 tasks x 8 trials"",""tools"":""domain customer-service tools"",""version"":""tau3 airline domain""}"
tau3_retail,tau3-bench Retail,Tool Use,% passed (8-trial average),912.0,https://github.com/sierra-research/tau2-bench,"{""higher_is_better"":true,""judge"":""state-based task success"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""114 tasks x 8 trials; extra simulator prompt and GPT-5.2-low simulator."",""range"":[0,100],""sampling"":""114 tasks x 8 trials"",""tools"":""domain customer-service tools"",""version"":""tau3 retail domain""}"
tau3_telecom,tau3-bench Telecom,Tool Use,% passed (8-trial average),912.0,https://github.com/sierra-research/tau2-bench,"{""higher_is_better"":true,""judge"":""state-based task success"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""114 tasks x 8 trials; extra simulator prompt and GPT-5.2-low simulator."",""range"":[0,100],""sampling"":""114 tasks x 8 trials"",""tools"":""domain customer-service tools"",""version"":""tau3 telecom domain""}"
ioi_2025,IOI 2025,Coding,contest points (0-600),6.0,https://ioi2025.bo/,"{""higher_is_better"":true,""metric_type"":""score"",""multimodal_input"":false,""notes"":""Six official contest problems; score is points, not percent."",""range"":[0,600],""sampling"":""pass@1"",""tools"":""code execution"",""version"":""International Olympiad in Informatics 2025""}"
aa_omniscience_accuracy,AA Omniscience Accuracy,Factuality,% correct,60000.0,https://artificialanalysis.ai/articles/aa-omniscience-knowledge-hallucination-benchmark,"{""higher_is_better"":true,""judge"":""official answer and abstention classification"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Count is 6,000 questions x 10 repeats. Accuracy=c/(c+p+i+a)."",""range"":[0,100],""sampling"":""10-run average"",""tools"":""none"",""version"":""AA-Omniscience 6,000-question benchmark""}"
aa_omniscience_non_hallucination,AA Omniscience Non-Hallucination,Hallucination,% non-hallucination,60000.0,https://artificialanalysis.ai/articles/aa-omniscience-knowledge-hallucination-benchmark,"{""higher_is_better"":true,""judge"":""official answer and abstention classification"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Count is 6,000 questions x 10 repeats. Non-hallucination=100-i/(p+i+a)."",""range"":[0,100],""sampling"":""10-run average"",""tools"":""none"",""version"":""AA-Omniscience 6,000-question benchmark""}"
ruler_1m,RULER 1M,Long Context,%,,https://github.com/NVIDIA/RULER,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official RULER aggregate at the exact 1M context length."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""RULER at 1M tokens""}"
wmt24pp,WMT24++,Multilingual,XCOMET-XXL,54890.0,https://huggingface.co/datasets/google/wmt24pp,"{""higher_is_better"":true,""judge"":""XCOMET-XXL"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""998 English paragraphs x 55 target languages = 54,890 translations."",""range"":[0,100],""sampling"":""54,890 translations"",""tools"":""none"",""version"":""WMT24++ English-to-55-language evaluation""}"
agieval_en,AGIEval English,Reasoning & Knowledge,% exact match,,https://github.com/microsoft/AGIEval,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""English tasks evaluated with benchmark-specific 3-shot or 5-shot CoT."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""AGIEval English aggregate""}"
math_test,MATH Test,Math,% exact match,5000.0,https://github.com/hendrycks/math,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Minerva 4-shot exact-match setting; distinct from full 12,500-example MATH."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""MATH test split (5,000 problems)""}"
mbpp_sanitized,MBPP Sanitized,Coding,sampled pass@1 %,13664.0,https://github.com/evalplus/evalplus,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Count is 427 problems x 32 samples = 13,664 model generations."",""range"":[0,100],""sampling"":""427 problems x 32 samples"",""tools"":""none"",""version"":""EvalPlus MBPP sanitized 427-problem set""}"
openbookqa,OpenBookQA,Knowledge,% normalized accuracy,500.0,https://github.com/allenai/OpenBookQA,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official 500-question test split, zero-shot normalized accuracy."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""OpenBookQA test split""}"
piqa,PIQA,Reasoning,% normalized accuracy,1838.0,https://yonatanbisk.com/piqa/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official 1,838-question validation split, zero-shot normalized accuracy."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""PIQA validation split""}"
winogrande,WinoGrande,Reasoning,% accuracy,1267.0,https://winogrande.allenai.org/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official 1,267-question validation split, 5-shot accuracy."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""WinoGrande validation split""}"
race,RACE,Reasoning & Knowledge,% accuracy,4934.0,https://www.cs.cmu.edu/~glai1/data/race/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official 4,934-question test split, zero-shot accuracy."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""RACE test split""}"
ruler_64k,RULER 64K,Long Context,%,,https://github.com/NVIDIA/RULER,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official RULER aggregate at the exact 64K context length."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""RULER at 64K tokens""}"
ruler_256k,RULER 256K,Long Context,%,,https://github.com/NVIDIA/RULER,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official RULER aggregate at the exact 256K context length."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""RULER at 256K tokens""}"
ruler_512k,RULER 512K,Long Context,%,,https://github.com/NVIDIA/RULER,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official RULER aggregate at the exact 512K context length."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""RULER at 512K tokens""}"
imo_proofbench_advanced,IMO-ProofBench Advanced,Math,% of 210 points,30.0,https://github.com/google-deepmind/superhuman/tree/main/imobench,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Thirty proof problems worth seven points each; 210 points total."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""IMO-ProofBench Advanced 30-problem split""}"
putnam_2025,Putnam 2025,Math,% of 120 points,12.0,https://maa.org/math-competitions/william-lowell-putnam-mathematical-competition,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Twelve proof problems worth ten points each; 120 points total."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""William Lowell Putnam Mathematical Competition 2025""}"
internal_research_debugging,Internal Research Debugging Evaluation,AI Self-Improvement,% tasks passed,,https://openai.com/index/gpt-5-6/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Named internal evaluation; task count not disclosed. The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""benchmark-specified"",""version"":""Internal Research Debugging Evaluation""}"
kernelgen_1p,KernelGen 1P,AI Self-Improvement,% reward,,https://openai.com/index/gpt-5-6/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""benchmark-specified"",""version"":""KernelGen 1P""}"
nanogpt_self_improvement,NanoGPT,AI Self-Improvement,% normalized reward,,https://openai.com/index/gpt-5-6/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""benchmark-specified"",""version"":""NanoGPT""}"
posttrain_bench_lite,PostTrainBench Lite,AI Self-Improvement,% score,,https://openai.com/index/gpt-5-6/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Distinct Lite variant from the existing PostTrainBench row. The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""benchmark-specified"",""version"":""PostTrainBench Lite""}"
rsi_index,RSI Index,AI Self-Improvement,index score,,https://openai.com/index/gpt-5-6/,"{""higher_is_better"":true,""metric_type"":""index"",""multimodal_input"":false,""notes"":""The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""benchmark-specified"",""version"":""RSI Index""}"
aa_coding_agent_index_v1_1,Artificial Analysis Coding Agent Index v1.1,Agentic Coding,index score,,https://artificialanalysis.ai/methodology/coding-agent,"{""higher_is_better"":true,""metric_type"":""index"",""multimodal_input"":false,""notes"":""The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""benchmark-specified"",""version"":""Artificial Analysis Coding Agent Index v1.1""}"
aav_capsid_packaging_prediction,AAV Capsid Packaging Prediction,Biology,Spearman correlation,,https://deploymentsafety.openai.com/gpt-5-6/aav-capsid-packaging-prediction,"{""higher_is_better"":true,""metric_type"":""correlation"",""multimodal_input"":false,""notes"":""The official source does not disclose a stable public task count."",""range"":[-1,1],""sampling"":""pass@1"",""tools"":""none"",""version"":""AAV Capsid Packaging Prediction""}"
dna_tf_binding_design,DNA sequence design for transcription factor binding,Biology,pass@1 (%),,https://deploymentsafety.openai.com/gpt-5-6/dna-sequence-design-for-transcription-factor-binding,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""DNA sequence design for transcription factor binding""}"
hard_negative_protein_binding,Hard-negative protein binding prediction,Biology,pass@4 (%),,https://deploymentsafety.openai.com/gpt-5-6/hard-negative-protein-binding-prediction,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@4"",""tools"":""none"",""version"":""Hard-negative protein binding prediction""}"
benchcad,BenchCAD,Computer Use,% tasks passed,,https://openai.com/index/gpt-5-6/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""BenchCAD""}"
exploitgym,ExploitGym,Cyber,% tasks passed,,https://openai.com/index/gpt-5-6/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""benchmark-specified"",""version"":""ExploitGym""}"
frontiercyber_easy,FrontierCyber Easy,Cyber,success rate (%),,https://deploymentsafety.openai.com/gpt-5-6,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Irregular agent harness; exact per-model denominators are stored on scores. The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""benchmark-specified"",""version"":""FrontierCyber Easy""}"
frontiercyber_elite,FrontierCyber Elite,Cyber,success rate (%),,https://deploymentsafety.openai.com/gpt-5-6,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Irregular agent harness; exact per-model denominators are stored on scores. The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""benchmark-specified"",""version"":""FrontierCyber Elite""}"
frontiercyber_hard,FrontierCyber Hard,Cyber,success rate (%),,https://deploymentsafety.openai.com/gpt-5-6,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Irregular agent harness; exact per-model denominators are stored on scores. The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""benchmark-specified"",""version"":""FrontierCyber Hard""}"
frontiercyber_medium,FrontierCyber Medium,Cyber,success rate (%),,https://deploymentsafety.openai.com/gpt-5-6,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Irregular agent harness; exact per-model denominators are stored on scores. The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""benchmark-specified"",""version"":""FrontierCyber Medium""}"
sec_bench_pro,SEC-Bench Pro,Cyber,% tasks passed,,https://openai.com/index/gpt-5-6/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""SEC-Bench Pro""}"
first_person_fairness_harm_overall,First-Person Fairness Evaluation,Fairness,harm_overall (%),,https://deploymentsafety.openai.com/gpt-5-6/first-person-fairness-evaluation,"{""higher_is_better"":false,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Expected male-versus-female biased-answer difference based on evaluation performance divided by 10; lower is better. The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""First-Person Fairness Evaluation""}"
big_finance_bench,Big Finance Bench,Finance,% tasks passed,,https://openai.com/index/gpt-5-6/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Distinct from the existing Finance Bench row. The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""Big Finance Bench""}"
healthbench_length_adjusted,HealthBench (length-adjusted),Health,length-adjusted score,5000.0,https://huggingface.co/datasets/openai/healthbench,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":"""",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""HealthBench (length-adjusted)""}"
healthbench_consensus_length_adjusted,HealthBench Consensus (length-adjusted),Health,length-adjusted score,3671.0,https://huggingface.co/datasets/openai/healthbench,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":"""",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""HealthBench Consensus (length-adjusted)""}"
healthbench_hard_length_adjusted,HealthBench Hard (length-adjusted),Health,length-adjusted score,1000.0,https://huggingface.co/datasets/openai/healthbench,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":"""",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""HealthBench Hard (length-adjusted)""}"
healthbench_professional,HealthBench Professional,Health,unadjusted score,,https://deploymentsafety.openai.com/gpt-5-6/healthbench,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""HealthBench Professional""}"
healthbench_professional_length_adjusted,HealthBench Professional (length-adjusted),Health,length-adjusted score,,https://deploymentsafety.openai.com/gpt-5-6/healthbench,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""HealthBench Professional (length-adjusted)""}"
management_consulting_tasks_internal,Management Consulting Tasks (Internal),Knowledge,% tasks passed,,https://openai.com/index/gpt-5-6/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Named internal evaluation; task count not disclosed. The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""Management Consulting Tasks (Internal)""}"
graphwalks_bfs_1m,GraphWalks BFS 1M,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact-length BFS point under the immutable official GraphWalks definition after the 2026-02-27 parents ground-truth and BFS prompt fixes. The deterministic judge computes set-overlap F1. Task counts remain observation-specific; no benchmark-wide count is invented."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; BFS exact 1M point""}"
graphwalks_bfs_256k,GraphWalks BFS 256K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact-length BFS point under the immutable official GraphWalks definition after the 2026-02-27 parents ground-truth and BFS prompt fixes. The deterministic judge computes set-overlap F1. Task counts remain observation-specific; no benchmark-wide count is invented."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; BFS exact 256K point""}"
mrcr_v2_8needle_256k_512k,"OpenAI MRCR v2 (8-needle, 256K-512K)",Long Context,% correct,200.0,https://huggingface.co/datasets/openai/mrcr,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":"""",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""OpenAI MRCR v2 (8-needle, 256K-512K)""}"
frontiermath_tier_1_3_v2,FrontierMath Tier 1-3 v2,Math,% correct,,https://epoch.ai/benchmarks/frontiermath,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Version 2 is distinct from the existing FrontierMath row. The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""FrontierMath Tier 1-3 v2""}"
frontiermath_tier_4_v2,FrontierMath Tier 4 v2,Math,% correct,,https://epoch.ai/benchmarks/frontiermath,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Version 2 is distinct from the existing FrontierMath Tier 4 row. The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""FrontierMath Tier 4 v2""}"
gdp_pdf,GDP.pdf,Multimodal,strict task pass rate (%),100.0,https://raw.githubusercontent.com/surge-ai/gdp-pdf/7a72a514a6ab19c90babb00adc817e4ae86b9c1b/README.md,"{""dataset_split"":""test"",""harness"":""Surge GDP.pdf Inspect AI harness"",""higher_is_better"":true,""judge"":""Gemini 3.5 Flash rubric judge"",""metric_type"":""all_pass_pct"",""multimodal_input"":true,""notes"":""Headline all_pass/mean: a task passes only when every rubric criterion is satisfied. The repository and HF size endpoint independently lock 100 tasks."",""range"":[0,100],""sampling"":""pass@1; published leaderboard uses five epochs"",""tools"":""none (per-benchmark override)"",""version"":""GDP.pdf 100-task held-out test set""}"
arc_agi_3,ARC-AGI-3,Reasoning,% tasks passed,,https://arcprize.org/arc-agi/3/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""ARC-AGI-3""}"
genebench_pro,GeneBench Pro,Science,% tasks passed,,https://openai.com/index/gpt-5-6/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Distinct professional variant from the existing GeneBench row. The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""GeneBench Pro""}"
lifescibench,LifeSciBench,Science,% tasks passed,,https://openai.com/index/gpt-5-6/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""LifeSciBench""}"
medchembench_internal,MedChemBench (Internal),Science,% tasks passed,,https://openai.com/index/gpt-5-6/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Named internal evaluation; task count not disclosed. The official source does not disclose a stable public task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""MedChemBench (Internal)""}"
long_form_virology_task_1_sequence_design,Long-form virology: Task 1 / Sequence design,Biology,score (0-1),,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""score"",""multimodal_input"":false,""notes"":""Anthropic internal automated CB-1 evaluation; exact task count is not disclosed."",""range"":[0,1],""tools"":""none"",""version"":""Long-form virology: Task 1 / Sequence design""}"
long_form_virology_task_1_protocol_design,Long-form virology: Task 1 / Protocol design,Biology,score (0-1),,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""score"",""multimodal_input"":false,""notes"":""Anthropic internal automated CB-1 evaluation; exact task count is not disclosed."",""range"":[0,1],""tools"":""none"",""version"":""Long-form virology: Task 1 / Protocol design""}"
long_form_virology_task_1_end_to_end,Long-form virology: Task 1 / End-to-end,Biology,score (0-1),,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""score"",""multimodal_input"":false,""notes"":""Anthropic internal automated CB-1 evaluation; exact task count is not disclosed."",""range"":[0,1],""tools"":""none"",""version"":""Long-form virology: Task 1 / End-to-end""}"
long_form_virology_task_2_sequence_design,Long-form virology: Task 2 / Sequence design,Biology,score (0-1),,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""score"",""multimodal_input"":false,""notes"":""Anthropic internal automated CB-1 evaluation; exact task count is not disclosed."",""range"":[0,1],""tools"":""none"",""version"":""Long-form virology: Task 2 / Sequence design""}"
long_form_virology_task_2_protocol_design,Long-form virology: Task 2 / Protocol design,Biology,score (0-1),,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""score"",""multimodal_input"":false,""notes"":""Anthropic internal automated CB-1 evaluation; exact task count is not disclosed."",""range"":[0,1],""tools"":""none"",""version"":""Long-form virology: Task 2 / Protocol design""}"
long_form_virology_task_2_end_to_end,Long-form virology: Task 2 / End-to-end,Biology,score (0-1),,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""score"",""multimodal_input"":false,""notes"":""Anthropic internal automated CB-1 evaluation; exact task count is not disclosed."",""range"":[0,1],""tools"":""none"",""version"":""Long-form virology: Task 2 / End-to-end""}"
ai_rd_kernel_best_speedup,AI R&D Kernel Best Speedup,Agentic Science,speedup (x),,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""ratio"",""multimodal_input"":false,""notes"":""Anthropic internal AI R&D rule-out evaluation; inherited protocol is cited to the Mythos 5 System Card."",""range"":[0,null],""tools"":""none"",""version"":""AI R&D Kernel Best Speedup""}"
ai_rd_time_series_forecasting_mse,AI R&D Time-Series Forecasting,Agentic Science,MSE,,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":false,""metric_type"":""error"",""multimodal_input"":false,""notes"":""Anthropic internal AI R&D rule-out evaluation; inherited protocol is cited to the Mythos 5 System Card."",""range"":[0,null],""tools"":""none"",""version"":""AI R&D Time-Series Forecasting""}"
ai_rd_llm_training_speedup,AI R&D LLM Training Speedup,Agentic Science,speedup (x),,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""ratio"",""multimodal_input"":false,""notes"":""Anthropic internal AI R&D rule-out evaluation; inherited protocol is cited to the Mythos 5 System Card."",""range"":[0,null],""tools"":""none"",""version"":""AI R&D LLM Training Speedup""}"
ai_rd_quadruped_rl,AI R&D Quadruped RL,Agentic Science,score,,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""score"",""multimodal_input"":false,""notes"":""Anthropic internal AI R&D rule-out evaluation; inherited protocol is cited to the Mythos 5 System Card."",""range"":[null,null],""tools"":""none"",""version"":""AI R&D Quadruped RL""}"
ai_rd_novel_compiler,AI R&D Novel Compiler,Agentic Science,% complex tests passed,,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Anthropic internal AI R&D rule-out evaluation; inherited protocol is cited to the Mythos 5 System Card."",""range"":[0,100],""tools"":""none"",""version"":""AI R&D Novel Compiler""}"
oss_fuzz_control_flow_hijack_count,OSS-Fuzz Exploit Primitive: Control-Flow Hijack,Cyber,targets reaching grade 1.0,830.0,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""count"",""multimodal_input"":false,""notes"":""Internal evaluation over about 830 OSS-Fuzz entry points from 228 projects."",""range"":[0,830],""tools"":""benchmark-specified"",""version"":""OSS-Fuzz Exploit Primitive: Control-Flow Hijack""}"
oss_fuzz_any_progress,OSS-Fuzz Exploit Primitive: Any Progress,Cyber,% targets with grade > 0,830.0,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Internal evaluation over about 830 OSS-Fuzz entry points from 228 projects."",""range"":[0,100],""tools"":""benchmark-specified"",""version"":""OSS-Fuzz Exploit Primitive: Any Progress""}"
firefox_147_exploit_development_working_exploit,Firefox 147 Exploit Development: Working Exploit,Cyber,% trials with working exploit,250.0,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""50 crash categories x 5 trials; security mitigations disabled."",""range"":[0,100],""tools"":""benchmark-specified"",""version"":""Firefox 147 Exploit Development: Working Exploit""}"
firefox_147_exploit_development_any_success,Firefox 147 Exploit Development: Any Success,Cyber,% trials with grade >= 0.5,250.0,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""50 crash categories x 5 trials; security mitigations disabled."",""range"":[0,100],""tools"":""benchmark-specified"",""version"":""Firefox 147 Exploit Development: Any Success""}"
frontiercode_main_v1,FrontierCode Main v1,Agentic Coding,score (%),100.0,https://cognition.com/blog/frontier-code,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The plotted values match FrontierCode Main (100 hardest tasks), not Extended or Diamond."",""range"":[0,100],""tools"":""benchmark-specified"",""version"":""FrontierCode Main v1""}"
automation_bench_private_heldout,AutomationBench Private Held-Out,Agentic,% tasks passed,,https://zapier.com/benchmarks,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Distinct private held-out leaderboard set; do not merge into the current public 600-task row."",""range"":[0,100],""tools"":""benchmark-specified"",""version"":""AutomationBench Private Held-Out""}"
legal_agent_benchmark_public,Legal Agent Benchmark: Full Public Set,Legal,all-pass rate (%),1235.0,https://www.harvey.ai/blog/introducing-harveys-legal-agent-benchmark,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""1,235 tested tasks after 16 pre-test exclusions from 1,251."",""range"":[0,100],""tools"":""benchmark-specified"",""version"":""Legal Agent Benchmark: Full Public Set""}"
legal_agent_benchmark_harvey_held_out,Legal Agent Benchmark: Harvey Held-Out,Legal,all-pass rate (%),,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Separate proprietary held-out set; task count is not disclosed."",""range"":[0,100],""tools"":""benchmark-specified"",""version"":""Legal Agent Benchmark: Harvey Held-Out""}"
cursorbench_3_1,CursorBench 3.1,Agentic Coding,score (%),,https://cursor.com/cursorbench,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The June 30 System Card predates CursorBench 3.2 (July 8); preserve it as CursorBench 3.1."",""range"":[0,100],""tools"":""benchmark-specified"",""version"":""CursorBench 3.1""}"
arxivmath_2026_04_05,ArxivMath April-May 2026,Math,accuracy (%),81.0,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact monthly releases: April 2026 (41) plus May 2026 (40)."",""range"":[0,100],""tools"":""none"",""version"":""ArxivMath April-May 2026""}"
gdp_pdf_mean_criteria_pass_rate,GDP.pdf Mean Criteria Pass Rate,Multimodal,mean criteria pass rate (%),100.0,https://surgehq.ai/benchmarks/gdp-pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Distinct from the current gdp_pdf '% tasks passed' row."",""range"":[0,100],""tools"":""none"",""version"":""GDP.pdf Mean Criteria Pass Rate""}"
chartmuseum,ChartMuseum,Multimodal,accuracy (%),1162.0,https://github.com/Liyan06/ChartMuseum,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Full 1,162-question benchmark."",""range"":[0,100],""tools"":""none"",""version"":""ChartMuseum""}"
real_world_finance_v2_elo,Real-World Finance v2,Finance,Elo rating,294.0,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""elo"",""multimodal_input"":false,""notes"":""Anthropic internal 294-task quantitative-finance evaluation."",""range"":[null,null],""tools"":""benchmark-specified"",""version"":""Real-World Finance v2""}"
global_mmlu,Global MMLU,Multilingual,average accuracy (%),589764.0,https://huggingface.co/datasets/CohereLabs/Global-MMLU,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Full test split across 42 languages."",""range"":[0,100],""tools"":""none"",""version"":""Global MMLU""}"
milu,MILU,Multilingual,average accuracy (%),79617.0,https://huggingface.co/datasets/ai4bharat/MILU,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Full 79,617-question test benchmark across 11 languages."",""range"":[0,100],""tools"":""none"",""version"":""MILU""}"
include_base_44,INCLUDE-base-44,Multilingual,average accuracy (%),22637.0,https://huggingface.co/datasets/CohereLabs/include-base-44,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Full 44-language base benchmark."",""range"":[0,100],""tools"":""none"",""version"":""INCLUDE-base-44""}"
biomysterybench_human_solvable,BioMysteryBench / Human solvable,Biology,score (0-1),,https://www.anthropic.com/research/Evaluating-Claude-For-Bioinformatics-With-BioMysteryBench,"{""higher_is_better"":true,""metric_type"":""score"",""multimodal_input"":false,""notes"":""Official BioMysteryBench subset; subset count is not disclosed in the System Card."",""range"":[0,1],""tools"":""none"",""version"":""BioMysteryBench / Human solvable""}"
biomysterybench_human_difficult,BioMysteryBench / Human difficult,Biology,score (0-1),,https://www.anthropic.com/research/Evaluating-Claude-For-Bioinformatics-With-BioMysteryBench,"{""higher_is_better"":true,""metric_type"":""score"",""multimodal_input"":false,""notes"":""Official BioMysteryBench subset; subset count is not disclosed in the System Card."",""range"":[0,1],""tools"":""none"",""version"":""BioMysteryBench / Human difficult""}"
spatialbench_verified,SpatialBench Verified,Biology,score (0-1),115.0,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""score"",""multimodal_input"":false,""notes"":""LatchBio externally validated spatial-transcriptomics benchmark."",""range"":[0,1],""tools"":""none"",""version"":""SpatialBench Verified""}"
singlecellbench,SingleCellBench,Biology,score (0-1),195.0,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""score"",""multimodal_input"":false,""notes"":""LatchBio single-cell RNA sequencing benchmark."",""range"":[0,1],""tools"":""none"",""version"":""SingleCellBench""}"
structural_biology_open_ended_internal,Structural biology open-ended,Biology,score (0-1),,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""score"",""multimodal_input"":false,""notes"":""Anthropic internal open-ended structural biology evaluation."",""range"":[0,1],""tools"":""none"",""version"":""Structural biology open-ended""}"
proteingym_hard,ProteinGym Hard,Biology,score (0-1),,https://proteingym.org/,"{""higher_is_better"":true,""metric_type"":""score"",""multimodal_input"":false,""notes"":""Exact hard split count is not disclosed in the System Card."",""range"":[0,1],""tools"":""none"",""version"":""ProteinGym Hard""}"
organic_chemistry_internal,Organic chemistry,Science,score (0-1),,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""score"",""multimodal_input"":false,""notes"":""Anthropic internal organic chemistry evaluation."",""range"":[0,1],""tools"":""none"",""version"":""Organic chemistry""}"
protocol_troubleshooting_internal,Protocol troubleshooting,Biology,score (0-1),,https://www.anthropic.com/claude-sonnet-5-system-card,"{""higher_is_better"":true,""metric_type"":""score"",""multimodal_input"":false,""notes"":""Anthropic internal molecular-biology protocol troubleshooting evaluation."",""range"":[0,1],""tools"":""none"",""version"":""Protocol troubleshooting""}"
ai_rd_llm_training_hard_speedup,AI R&D LLM Training Hard Speedup,Agentic Science,speedup (x),,https://www.anthropic.com/claude-opus-5-system-card,"{""higher_is_better"":true,""metric_type"":""ratio"",""multimodal_input"":false,""notes"":""Separate hard variant from the existing easy LLM-training speedup row."",""range"":[0,null],""tools"":""benchmark-specified"",""version"":""AI R&D LLM Training hard variant""}"
arxivmath_2026_06_no_tools,ArxivMath June 2026 (No Tools),Math,accuracy (%),49.0,https://matharena.ai/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Distinct June 2026 release; four runs per problem for Anthropic internal values."",""range"":[0,100],""tools"":""none"",""version"":""ArxivMath June 2026""}"
arxivmath_2026_06_with_tools,ArxivMath June 2026 (With Tools),Math,accuracy (%),49.0,https://matharena.ai/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Distinct June 2026 release; four runs per problem."",""range"":[0,100],""tools"":""benchmark-specified"",""version"":""ArxivMath June 2026""}"
biomysterybench_human_difficult_revised_2026_07,BioMysteryBench Human Difficult (Revised July 2026),Biology,score (0-1),,https://www.anthropic.com/claude-opus-5-system-card,"{""higher_is_better"":true,""metric_type"":""score"",""multimodal_input"":false,""notes"":""Distinct revised subset after removal of 6 Human Difficult problems."",""range"":[0,1],""tools"":""benchmark-specified"",""version"":""BioMysteryBench revised July 2026""}"
biomysterybench_human_solvable_revised_2026_07,BioMysteryBench Human Solvable (Revised July 2026),Biology,score (0-1),,https://www.anthropic.com/claude-opus-5-system-card,"{""higher_is_better"":true,""metric_type"":""score"",""multimodal_input"":false,""notes"":""Distinct revised subset after removal of 3 Human Solvable problems."",""range"":[0,1],""tools"":""benchmark-specified"",""version"":""BioMysteryBench revised July 2026""}"
chartography_no_tools,Chartography (No Tools),Multimodal,score (%),100.0,https://www.surgehq.ai/blog/chartography,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""100 specialized chart types; five runs."",""range"":[0,100],""tools"":""none"",""version"":""Chartography""}"
chartography_with_tools,Chartography (With Tools),Multimodal,score (%),100.0,https://www.surgehq.ai/blog/chartography,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""100 specialized chart types; five runs."",""range"":[0,100],""tools"":""benchmark-specified"",""version"":""Chartography""}"
cursorbench_3_2,CursorBench 3.2,Agentic Coding,score (%),,https://cursor.com/cursorbench,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Ambiguous multi-file tasks from real Cursor sessions; task count is not disclosed."",""range"":[0,100],""tools"":""Cursor coding agent"",""version"":""CursorBench 3.2""}"
draco,DRACO,Agentic Data Analysis,normalized score (%),100.0,https://www.anthropic.com/claude-opus-5-system-card,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""100 curated tasks; four grading categories; five independent grading runs."",""range"":[0,100],""tools"":""web search, web fetch, programmatic calls, code execution"",""version"":""DRACO""}"
frontierbench_v0_1,FrontierBench v0.1,Agentic,% tasks completed,74.0,https://github.com/harbor-framework/frontier-bench,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""74 professional terminal tasks. Harness differences are retained in reported_setting."",""range"":[0,100],""tools"":""terminal agent"",""version"":""FrontierBench v0.1""}"
frontiercode_extended_v1_1,FrontierCode v1.1 Extended,Agentic Coding,score (%),150.0,https://cognition.com/blog/frontier-code-1.1,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Private full 150-task set; five runs per available effort."",""range"":[0,100],""tools"":""coding agent"",""version"":""FrontierCode v1.1 Extended""}"
frontiercode_main_v1_1,FrontierCode v1.1 Main,Agentic Coding,score (%),100.0,https://cognition.com/blog/frontier-code-1.1,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Private 100-task Main subset; five runs per available effort."",""range"":[0,100],""tools"":""coding agent"",""version"":""FrontierCode v1.1 Main""}"
gmmlu,GMMLU,Multilingual,average accuracy (%),,https://www.anthropic.com/claude-opus-5-system-card,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""42-language evaluation; the System Card does not disclose a stable task count."",""range"":[0,100],""tools"":""none"",""version"":""GMMLU""}"
organic_chemistry_v2_internal,Organic Chemistry V2,Science,score (0-1),,https://www.anthropic.com/claude-opus-5-system-card,"{""higher_is_better"":true,""metric_type"":""score"",""multimodal_input"":false,""notes"":""Versioned separately from the earlier internal Organic Chemistry evaluation."",""range"":[0,1],""tools"":""benchmark-specified"",""version"":""Anthropic internal Organic Chemistry V2""}"
protein_design_internal,Protein Design,Biology,score (0-1),,https://www.anthropic.com/claude-opus-5-system-card,"{""higher_is_better"":true,""metric_type"":""score"",""multimodal_input"":false,""notes"":""Combined constraint-satisfaction, folding-confidence, and novelty score."",""range"":[0,1],""tools"":""none"",""version"":""Anthropic internal Protein Design""}"
protocol_understanding_internal,Protocol Understanding (Benchling),Biology,score (0-1),,https://www.anthropic.com/claude-opus-5-system-card,"{""higher_is_better"":true,""metric_type"":""score"",""multimodal_input"":false,""notes"":""Distinct from Protocol Troubleshooting."",""range"":[0,1],""tools"":""bash, file editor, web search"",""version"":""Protocol Understanding (Benchling)""}"
riemannbench_no_tools,RiemannBench (No Tools),Math,score (%),25.0,https://arxiv.org/abs/2604.06802,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Private 25-problem research-mathematics benchmark; mean over four attempts."",""range"":[0,100],""tools"":""none"",""version"":""RiemannBench corrected references/grading""}"
riemannbench_with_tools,RiemannBench (With Tools),Math,score (%),25.0,https://arxiv.org/abs/2604.06802,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Private 25-problem research-mathematics benchmark; mean over four attempts."",""range"":[0,100],""tools"":""benchmark-specified"",""version"":""RiemannBench corrected references/grading""}"
cursorbench_3_0,CursorBench 3.0,Agentic Coding,score (%),,https://cursor.com/resources/Composer2.pdf,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The Composer 2 report calls this CursorBench-3. Its 2026-03-25 evaluation predates CursorBench 3.1 (2026-05-19) and follows the 3.0 launch (2026-03-11), so the exact version identity is CursorBench 3.0. Task count is undisclosed."",""range"":[0,100],""tools"":""Cursor coding agent"",""version"":""CursorBench 3.0""}"
mrcr_v2_8needle_512k_1m,MRCR v2 8-Needle 512K-1M,Long Context,mean SequenceMatcher ratio (%),100.0,https://huggingface.co/datasets/openai/mrcr,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official dataset has 100 samples per context-length bin. Scoring uses the mean difflib SequenceMatcher ratio with the required hash prefix. Distinct from the all-bin 800-task aggregate."",""range"":[0,100],""tools"":""none"",""version"":""OpenAI MRCR; 8 needles; (524,288, 1,048,576] token bin""}"
mcpatlas_full_1000,"MCP-Atlas Full 1,000",Agentic,pass rate at >=0.75 claim coverage (%),1000.0,https://arxiv.org/abs/2602.00933v3,"{""harness"":""Scale AI MCP-Atlas agent harness and containerized scoring pipeline"",""higher_is_better"":true,""judge"":""claim-level 1/0.5/0 scoring; Gemini 3.1 Pro Preview primary judge with GPT-5.4 and Claude Opus 4.6 sensitivity judges"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Full set is 500 public plus 500 held-out private tasks. A task passes when mean claim coverage is at least 0.75."",""range"":[0,100],""sampling"":""pass@1 over all 1,000 tasks"",""tools"":""controlled target and distractor tools across 36 real MCP servers and 220 tools"",""version"":""MCP-Atlas full 1,000-task evaluation""}"
osworld_2_0_binary,OSWorld 2.0 Binary,Agentic,binary success (%),108.0,https://osworld-v2.xlang.ai/,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Binary success metric printed separately from OSWorld 2.0 partial-credit score over the 108-task release."",""range"":[0,100],""tools"":""computer-use environment"",""version"":""OSWorld 2.0 binary scoring""}"
webarena_verified_full,WebArena-Verified Full,Agentic,task success (%),812.0,https://github.com/ServiceNow/webarena-verified,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Full verified set, distinct from the 258-task Hard subset and from original WebArena."",""range"":[0,100],""tools"":""browser"",""version"":""WebArena-Verified full 812-task set""}"
gdpval_aa_v2_elo,GDPval-AA v2 Elo,Professional,Elo,220.0,https://artificialanalysis.ai/methodology/intelligence-benchmarking#gdpval-aa,"{""harness"":""Artificial Analysis Stirrup in a fresh E2B sandbox; 250-turn limit"",""higher_is_better"":true,""judge"":""blind pairwise panel of three frontier LLM judges; Bradley-Terry Elo"",""metric_type"":""elo"",""multimodal_input"":false,""notes"":""All 220 public OpenAI GDPval gold tasks across 44 occupations. The v2 Elo scale is anchored to human-expert deliverables at 1000."",""range"":null,""sampling"":""one agentic submission per task"",""tools"":""Web Fetch, Web Search, View Image, Code Exec, Finish, Abandon Task"",""version"":""GDPval-AA v2""}"
meta_internal_coding_bench,Meta Internal Coding Bench,Agentic Coding,% resolved (pass@1),440.0,https://research.meta.ai/static/muse-spark-1-2-methodology,"{""harness"":""Meta internal agentic harness; internet disabled"",""higher_is_better"":true,""judge"":""compile and unit-test verifier in dedicated grading containers"",""metric_type"":""pass_at_1_pct"",""multimodal_input"":false,""notes"":""Private 440-task benchmark derived from real internal pull requests."",""range"":[0,100],""sampling"":""two attempts per task; average task-level success rate"",""tools"":""internal agentic coding environment"",""version"":""Meta Internal Coding Bench, August 2026""}"
vibecodebench_v1_1_test,Vibe Code Bench v1.1 Test,Agentic Coding,mean per-application accuracy (%),50.0,https://www.vals.ai/benchmarks/vibe-code,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official benchmark has 50 validation and 50 held-out test tasks; leaderboard scores use the 50-task held-out test set. A workflow passes when at least 90% of its substeps succeed."",""range"":[0,100],""tools"":""agentic coding environment"",""version"":""Vibe Code Bench v1.1 held-out test set""}"
swe_atlas_codebase_qna,SWE-Atlas-QnA,Agentic Coding,task resolve rate (%),124.0,https://huggingface.co/datasets/SWE-Atlas/SWE-Atlas-QnA,"{""harness"":""SWE-Atlas-QnA official dataset and score-level agent harness"",""higher_is_better"":true,""judge"":""repository-grounded task-resolution evaluator"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Repository question answering, not patch generation; 124 tasks across 11 repositories. Agent harness remains score-level provenance."",""range"":[0,100],""sampling"":""unknown"",""tools"":""codebase reading and exploration"",""version"":""SWE-Atlas-QnA public 124-task split""}"
wmdp_bio_accuracy,WMDP-Bio,Safety Capability,accuracy (%),1273.0,https://wmdp.ai,"{""harness"":""WMDP official evaluation"",""higher_is_better"":true,""judge"":""multiple-choice exact accuracy"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Public Bio split has 1,273 questions; xAI capability rows are evaluated without safeguards."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none; no safeguards"",""version"":""WMDP-Bio current public release""}"
wmdp_chem_accuracy,WMDP-Chem,Safety Capability,accuracy (%),408.0,https://wmdp.ai,"{""harness"":""WMDP official evaluation"",""higher_is_better"":true,""judge"":""multiple-choice exact accuracy"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Public Chemistry split has 408 questions; xAI capability rows are evaluated without safeguards."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none; no safeguards"",""version"":""WMDP-Chem current public release""}"
lab_bench_protocolqa_accuracy,LAB-Bench ProtocolQA,Safety Capability,accuracy (%),108.0,https://github.com/Future-House/LAB-Bench,"{""harness"":""LAB-Bench official ProtocolQA harness"",""higher_is_better"":true,""judge"":""official multiple-choice exact-match accuracy"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The xAI cards report a distinct open-ended ProtocolQA adaptation. Those scores remain noncanonical score-level variants rather than redefining this benchmark id."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""LAB-Bench ProtocolQA official 108-question MCQ split""}"
ipho_2025_theory,IPhO 2025 Theory,Science,normalized theory marks (%),3.0,https://www.ipho2025.fr/official-questions-ipho-france-2025,"{""higher_is_better"":true,""judge"":""official partial-credit marking scheme"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""The organizer publishes three used theory problems. Google's scores average eight runs and use Gemini as judge, so those observations are retained as noncanonical settings."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""IPhO 2025 full theory examination: 3 used problems""}"
icho_2025_theory,IChO 2025 Theory,Science,normalized theory marks (%),9.0,https://www.icho-official.org/results/results.php?id=57&year=2025,"{""higher_is_better"":true,""judge"":""official partial-credit marking scheme"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Official theory marks are normalized to 100. Google's scores average eight runs and use Gemini as judge, so those observations are retained as noncanonical settings."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""IChO 2025 full theory examination: 9 problems""}"
blueprint_bench_2,Blueprint-Bench 2,Spatial Reasoning,normalized connectivity score (%),50.0,https://andonlabs.com/evals/blueprint-bench-2,"{""higher_is_better"":true,""judge"":""D4-invariant deterministic connectivity composite"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Composite weights: Jaccard 50%, degree 20%, density 10%, room count 10%, door count 5%, orientation 5%. Full v2 dataset/evaluator is not public, so reported cells are stored with matches_canonical=false."",""range"":[0,100],""sampling"":""leaderboard protocol"",""tools"":""persistent cross-apartment notepad"",""version"":""Blueprint-Bench 2: 50-apartment sequential evaluation""}"
gdm_mrcr_v2_8needle_upto_128k,GDM MRCR v2 8-Needle up to 128K,Long Context,mean strict MRCR score (%),484.0,https://github.com/google-deepmind/eval_hub/tree/master/eval_hub/mrcr_v2,"{""higher_is_better"":true,""judge"":""official hash check + SequenceMatcher strict scorer"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Google DeepMind MRCR v2, not OpenAI MRCR v2. Count 484 data rows in the official fixed CSV SHA 706a254439905ed6286d1184a03a018667e0ea0f49ad3312aedb8c27054ab0bd."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""mrcr_v2p1 8-needle upto_128K cumulative CSV""}"
gdm_mrcr_v2_8needle_1m,GDM MRCR v2 8-Needle at 1M,Long Context,mean strict MRCR score (%),,https://github.com/google-deepmind/eval_hub/tree/master/eval_hub/mrcr_v2,"{""higher_is_better"":true,""judge"":""official hash check + SequenceMatcher strict scorer"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Google DeepMind MRCR v2, not OpenAI MRCR v2. Official 1.5GB fixed object and scorer are public; the README does not publish the data-row count, so num_problems remains null."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""mrcr_v2p1 8-needle (524288,1048576] pointwise CSV""}"
mle_bench_partial30_avg_position_k2,MLE-Bench Partial 30 Average Position (k=2),Agentic Coding,Average Position Score (%),60.0,https://github.com/openai/mle-bench,"{""higher_is_better"":true,""judge"":""Kaggle private-leaderboard rank transformed by (N-r+1)/N"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""30 fixed competitions from experiments/splits/systemcard.txt and two independent runs each: 60 model episodes. Failed/missing submission scores zero; metric is not Any-Medal accuracy."",""range"":[0,100],""sampling"":""2 independent runs per competition"",""tools"":""interactive Bash, internet, isolated H100 sandbox"",""version"":""official systemcard Partial 30 split; k=2 runs""}"
infographicvqa,InfographicVQA,Vision,ANLS (%),3288.0,https://arxiv.org/abs/2104.12756,"{""higher_is_better"":true,""metric_type"":""anls"",""multimodal_input"":true,""notes"":""Official test split; ANLS uses normalized Levenshtein similarity."",""range"":[0,100],""tools"":""none"",""version"":""InfographicVQA test split""}"
covost2_xx_en_7lang_macro,CoVoST2 XX-to-English 7-Language Macro,Audio,macro CorpusBLEU,62325.0,https://huggingface.co/datasets/facebook/covost2,"{""higher_is_better"":true,""metric_type"":""bleu"",""multimodal_input"":true,""notes"":""Simple macro over seven direction-level CorpusBLEU scores; Gemma uses a transcribe-then-translate prompt."",""range"":[0,100],""tools"":""none"",""version"":""CoVoST2 ja/de/fr/es/it/ru/zh-CN to English macro""}"
loft_text_retrieval_128k,LOFT Text Retrieval at 128K,Long Context,Recall@k (%),,https://github.com/google-deepmind/loft,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Gemma reports an aggregate across LOFT text-retrieval datasets; exact included datasets and weighting are not published."",""range"":[0,100],""tools"":""none"",""version"":""LOFT text-retrieval aggregate at 128K""}"
graphwalks_lt128k_combined,GraphWalks Combined below 128K,Long Context,F1 (%),650.0,https://huggingface.co/datasets/openai/graphwalks,"{""higher_is_better"":true,""metric_type"":""f1"",""multimodal_input"":false,""notes"":""Combined score over 300 BFS and 350 parent-node rows; the report does not publish the aggregation weighting."",""range"":[0,100],""tools"":""none"",""version"":""GraphWalks BFS plus parent-node prompts below 128K""}"
covost2_ja_en,CoVoST2 ja to English,Audio,CorpusBLEU,684.0,https://huggingface.co/datasets/facebook/covost2,"{""higher_is_better"":true,""metric_type"":""bleu"",""multimodal_input"":true,""notes"":""Speech-to-English translation; Gemma uses a transcribe-then-translate prompt."",""range"":[0,100],""tools"":""none"",""version"":""CoVoST2 ja to English test direction""}"
covost2_de_en,CoVoST2 de to English,Audio,CorpusBLEU,13511.0,https://huggingface.co/datasets/facebook/covost2,"{""higher_is_better"":true,""metric_type"":""bleu"",""multimodal_input"":true,""notes"":""Speech-to-English translation; Gemma uses a transcribe-then-translate prompt."",""range"":[0,100],""tools"":""none"",""version"":""CoVoST2 de to English test direction""}"
covost2_fr_en,CoVoST2 fr to English,Audio,CorpusBLEU,14760.0,https://huggingface.co/datasets/facebook/covost2,"{""higher_is_better"":true,""metric_type"":""bleu"",""multimodal_input"":true,""notes"":""Speech-to-English translation; Gemma uses a transcribe-then-translate prompt."",""range"":[0,100],""tools"":""none"",""version"":""CoVoST2 fr to English test direction""}"
covost2_es_en,CoVoST2 es to English,Audio,CorpusBLEU,13221.0,https://huggingface.co/datasets/facebook/covost2,"{""higher_is_better"":true,""metric_type"":""bleu"",""multimodal_input"":true,""notes"":""Speech-to-English translation; Gemma uses a transcribe-then-translate prompt."",""range"":[0,100],""tools"":""none"",""version"":""CoVoST2 es to English test direction""}"
covost2_it_en,CoVoST2 it to English,Audio,CorpusBLEU,8951.0,https://huggingface.co/datasets/facebook/covost2,"{""higher_is_better"":true,""metric_type"":""bleu"",""multimodal_input"":true,""notes"":""Speech-to-English translation; Gemma uses a transcribe-then-translate prompt."",""range"":[0,100],""tools"":""none"",""version"":""CoVoST2 it to English test direction""}"
covost2_ru_en,CoVoST2 ru to English,Audio,CorpusBLEU,6300.0,https://huggingface.co/datasets/facebook/covost2,"{""higher_is_better"":true,""metric_type"":""bleu"",""multimodal_input"":true,""notes"":""Speech-to-English translation; Gemma uses a transcribe-then-translate prompt."",""range"":[0,100],""tools"":""none"",""version"":""CoVoST2 ru to English test direction""}"
covost2_zh_cn_en,CoVoST2 zh-CN to English,Audio,CorpusBLEU,4898.0,https://huggingface.co/datasets/facebook/covost2,"{""higher_is_better"":true,""metric_type"":""bleu"",""multimodal_input"":true,""notes"":""Speech-to-English translation; Gemma uses a transcribe-then-translate prompt."",""range"":[0,100],""tools"":""none"",""version"":""CoVoST2 zh-CN to English test direction""}"
fleurs_asr_en_us,FLEURS ASR en,Audio,WER,647.0,https://huggingface.co/datasets/google/fleurs,"{""higher_is_better"":false,""metric_type"":""wer"",""multimodal_input"":true,""notes"":""Automatic speech recognition transcription; lower is better."",""range"":null,""tools"":""none"",""version"":""FLEURS en test split""}"
fleurs_asr_ko_kr_cer,FLEURS ASR ko,Audio,CER,382.0,https://huggingface.co/datasets/google/fleurs,"{""higher_is_better"":false,""metric_type"":""cer"",""multimodal_input"":true,""notes"":""Automatic speech recognition transcription; lower is better."",""range"":null,""tools"":""none"",""version"":""FLEURS ko test split""}"
fleurs_asr_ja_jp_cer,FLEURS ASR ja,Audio,CER,650.0,https://huggingface.co/datasets/google/fleurs,"{""higher_is_better"":false,""metric_type"":""cer"",""multimodal_input"":true,""notes"":""Automatic speech recognition transcription; lower is better."",""range"":null,""tools"":""none"",""version"":""FLEURS ja test split""}"
fleurs_asr_de_de,FLEURS ASR de,Audio,WER,862.0,https://huggingface.co/datasets/google/fleurs,"{""higher_is_better"":false,""metric_type"":""wer"",""multimodal_input"":true,""notes"":""Automatic speech recognition transcription; lower is better."",""range"":null,""tools"":""none"",""version"":""FLEURS de test split""}"
fleurs_asr_fr_fr,FLEURS ASR fr,Audio,WER,676.0,https://huggingface.co/datasets/google/fleurs,"{""higher_is_better"":false,""metric_type"":""wer"",""multimodal_input"":true,""notes"":""Automatic speech recognition transcription; lower is better."",""range"":null,""tools"":""none"",""version"":""FLEURS fr test split""}"
fleurs_asr_hi_in,FLEURS ASR hi,Audio,WER,418.0,https://huggingface.co/datasets/google/fleurs,"{""higher_is_better"":false,""metric_type"":""wer"",""multimodal_input"":true,""notes"":""Automatic speech recognition transcription; lower is better."",""range"":null,""tools"":""none"",""version"":""FLEURS hi test split""}"
fleurs_asr_es_419,FLEURS ASR es,Audio,WER,908.0,https://huggingface.co/datasets/google/fleurs,"{""higher_is_better"":false,""metric_type"":""wer"",""multimodal_input"":true,""notes"":""Automatic speech recognition transcription; lower is better."",""range"":null,""tools"":""none"",""version"":""FLEURS es test split""}"
fleurs_asr_it_it,FLEURS ASR it,Audio,WER,865.0,https://huggingface.co/datasets/google/fleurs,"{""higher_is_better"":false,""metric_type"":""wer"",""multimodal_input"":true,""notes"":""Automatic speech recognition transcription; lower is better."",""range"":null,""tools"":""none"",""version"":""FLEURS it test split""}"
fleurs_asr_pt_br,FLEURS ASR pt-br,Audio,WER,919.0,https://huggingface.co/datasets/google/fleurs,"{""higher_is_better"":false,""metric_type"":""wer"",""multimodal_input"":true,""notes"":""Automatic speech recognition transcription; lower is better."",""range"":null,""tools"":""none"",""version"":""FLEURS pt-br test split""}"
fleurs_asr_ru_ru,FLEURS ASR ru,Audio,WER,775.0,https://huggingface.co/datasets/google/fleurs,"{""higher_is_better"":false,""metric_type"":""wer"",""multimodal_input"":true,""notes"":""Automatic speech recognition transcription; lower is better."",""range"":null,""tools"":""none"",""version"":""FLEURS ru test split""}"
fleurs_asr_ar_eg,FLEURS ASR ar,Audio,WER,428.0,https://huggingface.co/datasets/google/fleurs,"{""higher_is_better"":false,""metric_type"":""wer"",""multimodal_input"":true,""notes"":""Automatic speech recognition transcription; lower is better."",""range"":null,""tools"":""none"",""version"":""FLEURS ar test split""}"
fleurs_asr_zh_cn_cer,FLEURS ASR zh,Audio,CER,945.0,https://huggingface.co/datasets/google/fleurs,"{""higher_is_better"":false,""metric_type"":""cer"",""multimodal_input"":true,""notes"":""Automatic speech recognition transcription; lower is better."",""range"":null,""tools"":""none"",""version"":""FLEURS zh test split""}"
mtob_eng_kgv_half_book,MTOB English to Kalamang: Half Book,Long Context,chrF,100.0,https://github.com/lukemelas/mtob,"{""higher_is_better"":true,""metric_type"":""chrf"",""multimodal_input"":false,""notes"":""One hundred held-out translation sentence pairs."",""range"":null,""tools"":""none"",""version"":""MTOB eng-to-kgv with approximately 128K half-book context""}"
mtob_eng_kgv_full_book,MTOB English to Kalamang: Full Book,Long Context,chrF,100.0,https://github.com/lukemelas/mtob,"{""higher_is_better"":true,""metric_type"":""chrf"",""multimodal_input"":false,""notes"":""One hundred held-out translation sentence pairs."",""range"":null,""tools"":""none"",""version"":""MTOB eng-to-kgv with approximately 256K full-book context""}"
mtob_kgv_eng_half_book,MTOB Kalamang to English: Half Book,Long Context,chrF,100.0,https://github.com/lukemelas/mtob,"{""higher_is_better"":true,""metric_type"":""chrf"",""multimodal_input"":false,""notes"":""One hundred held-out translation sentence pairs."",""range"":null,""tools"":""none"",""version"":""MTOB kgv-to-eng with approximately 128K half-book context""}"
mtob_kgv_eng_full_book,MTOB Kalamang to English: Full Book,Long Context,chrF,100.0,https://github.com/lukemelas/mtob,"{""higher_is_better"":true,""metric_type"":""chrf"",""multimodal_input"":false,""notes"":""One hundred held-out translation sentence pairs."",""range"":null,""tools"":""none"",""version"":""MTOB kgv-to-eng with approximately 256K full-book context""}"
codeelo_rating,CodeElo,Coding,Elo rating,408.0,https://github.com/QwenLM/CodeElo,"{""higher_is_better"":true,""judge"":""Codeforces online judge"",""metric_type"":""elo"",""multimodal_input"":false,""notes"":""Eight samples per problem produce 3,264 judged submissions. The Google report does not disclose its replicate count, so its observations do not match canonical sampling."",""range"":null,""sampling"":""samples=8 (3264 submissions)"",""tools"":""none"",""version"":""CodeElo 408-problem Codeforces evaluation""}"
lbpp_v2_multilingual,LBPP v2 Multilingual,Coding,functional pass@1 (%),943.0,https://huggingface.co/datasets/CohereLabs/lbpp,"{""higher_is_better"":true,""judge"":""sandboxed unit tests"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The public multilingual configuration has 944 rows across Python, C++, Go, Java, JavaScript, and Rust; lbpp/python/042 is the mandatory canary and is excluded, leaving 943 scored tasks."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""LBPP v2 multilingual six-language test set after mandatory Python canary removal""}"
natural2code,Natural2Code (Google internal),Coding,% correct,,https://arxiv.org/abs/2608.00146,"{""higher_is_better"":true,""judge"":""Google internal evaluator (undisclosed)"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Internal/proprietary benchmark. Public task definition, count, and judge are unavailable; source observations remain valid provider-reported scores."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""Google internal Natural2Code evaluation""}"
pubmedqa_pqal_decision_accuracy,PubMedQA PQA-L Decision Accuracy,Knowledge/Medical,decision accuracy (%),500.0,https://github.com/pubmedqa/pubmedqa,"{""higher_is_better"":true,""judge"":""exact yes/no/maybe decision accuracy"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The official DiffusionGemma adapter maps the 500 PQA-L test IDs to generated categorical and long-form outputs."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""PubMedQA PQA-L official 500-item test set: yes/no/maybe decision""}"
pubmedqa_pqal_long_answer_bleu,PubMedQA PQA-L Long-Answer BLEU,Knowledge/Medical,corpus BLEU,500.0,https://github.com/pubmedqa/pubmedqa,"{""higher_is_better"":true,""judge"":""corpus BLEU against PQA-L long answers"",""metric_type"":""bleu"",""multimodal_input"":false,""notes"":""Corpus BLEU over the same 500 generated responses used for the decision metric by the official DiffusionGemma adapter."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""PubMedQA PQA-L official 500-item test set: generated long answers""}"
deep_swe_v1_0,DeepSWE v1.0,Agentic Coding,% resolved (pass@1),113.0,https://deepswe.datacurve.ai/,"{""harness"":""provider harnesses run by Artificial Analysis"",""higher_is_better"":true,""judge"":""behavioral and correctness verifiers"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Contamination-resistant repository issues; provider harness is score-level provenance. The xAI card does not disclose the trial count, so its scores are noncanonical variants."",""range"":[0,100],""sampling"":""4 trials; pass@1 per attempt"",""tools"":""agentic repository shell/editor"",""version"":""DeepSWE v1.0 113-task release""}"
apex_swe,APEX-SWE,Agentic Coding,pass@1 (%),200.0,https://www.mercor.com/apex/apex-swe-leaderboard/,"{""harness"":""Mercor APEX-SWE harness"",""higher_is_better"":true,""judge"":""Mercor integration and observability task verifier"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Integration and observability software-engineering tasks."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""agentic software-engineering environment"",""version"":""APEX-SWE 200-task release""}"
swe_marathon_v1_1,SWE-Marathon v1.1,Agentic Coding,"resolution rate (pass@1, %)",20.0,https://www.swe-marathon.org/,"{""harness"":""SWE-Marathon v1.1 harness"",""higher_is_better"":true,""judge"":""multi-layer resolution verification"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Ultra-long-horizon tasks with reward-hacking-resistant verification."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""agentic code execution"",""version"":""SWE-Marathon v1.1 full 20-task set""}"
falseclaimbench_accuracy,FalseClaimBench,Agent Reliability,fully true claim accuracy (%),,https://media.x.ai/v1/website/4p5-5184fdf9.pdf,"{""harness"":""SpaceXAI internal harness"",""higher_is_better"":true,""judge"":""claim-to-final-workspace-state verifier"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Internal task count is undisclosed; checks whether claimed work was actually performed."",""range"":[0,100],""sampling"":""unknown"",""tools"":""workspace editing and command tools"",""version"":""SpaceXAI internal FalseClaimBench""}"
eebench_v1_core_reward,EEBench v1 Core,Engineering,reward (%),13.0,https://eebench.org/,"{""harness"":""EEBench public leaderboard harness"",""higher_is_better"":true,""judge"":""physical correctness and functionality reward"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Public v1 core corpus contains 13 tasks; the leaderboard averages three trials."",""range"":[0,100],""sampling"":""3 trials"",""tools"":""electrical-engineering design tools"",""version"":""EEBench v1 core corpus""}"
parametric_cad_bench_combined,Parametric CAD Bench,Engineering,combined reward (%),100.0,https://www.gnucleus.ai/cad-bench,"{""harness"":""gNucleus CAD-Bench harness"",""higher_is_better"":true,""judge"":""combined executability and geometric reward"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official benchmark has 100 parametric CAD tasks."",""range"":[0,100],""sampling"":""unknown"",""tools"":""parametric CAD agent tools"",""version"":""Parametric CAD Bench combined score""}"
cadgenbench_generation_cad_score,CadGenBench Generation,Engineering,generation CAD score (%),49.0,https://github.com/huggingface/cadgenbench,"{""harness"":""CadGenBench generation harness"",""higher_is_better"":true,""judge"":""generation CAD score"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Generation split contains 49 public tasks."",""range"":[0,100],""sampling"":""unknown"",""tools"":""CAD generation agent tools"",""version"":""CadGenBench 49-task generation split""}"
spacexai_mts_accuracy,SpaceXAI MTS Eval,AI R&D,correct (%),29.0,https://media.x.ai/v1/website/4p5-5184fdf9.pdf,"{""harness"":""Grok Build"",""higher_is_better"":true,""judge"":""task correctness under fixed rollout budget"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The card explicitly identifies 29 internal model-development tasks; refused tasks are excluded where stated."",""range"":[0,100],""sampling"":""unknown"",""tools"":""Grok Build coding and research tools"",""version"":""SpaceXAI internal MTS 29-task suite""}"
single_turn_hallucination_rate,Single-Turn Hallucination,Factuality,hallucination rate (%),,https://media.x.ai/v1/website/4p5-5184fdf9.pdf,"{""harness"":""SpaceXAI internal harness"",""higher_is_better"":false,""judge"":""separate factual-support grader"",""metric_type"":""rate_pct"",""multimodal_input"":false,""notes"":""Internal task count is undisclosed; lower is better."",""range"":[0,100],""sampling"":""unknown"",""tools"":""none"",""version"":""SpaceXAI internal single-turn factuality suite""}"
hackerbench_v0_2_harmful_compliance,HackerBench v0.2 Harmful Compliance,Safety,harmful/dual-use compliance rate (%),,https://media.x.ai/v1/website/4p5-5184fdf9.pdf,"{""harness"":""HackerBench v0.2 with standard safeguards"",""higher_is_better"":false,""judge"":""harmful or dual-use compliance grader"",""metric_type"":""rate_pct"",""multimodal_input"":false,""notes"":""Internal task count is undisclosed; lower is better."",""range"":[0,100],""sampling"":""unknown"",""tools"":""constrained cyber task tools"",""version"":""HackerBench v0.2 harmful and dual-use split""}"
hackerbench_v0_2_benign_refusal,HackerBench v0.2 Benign Refusal,Safety,benign refusal rate (%),,https://media.x.ai/v1/website/4p5-5184fdf9.pdf,"{""harness"":""HackerBench v0.2 with standard safeguards"",""higher_is_better"":false,""judge"":""benign refusal grader"",""metric_type"":""rate_pct"",""multimodal_input"":false,""notes"":""Internal task count is undisclosed; lower is better."",""range"":[0,100],""sampling"":""unknown"",""tools"":""constrained cyber task tools"",""version"":""HackerBench v0.2 benign split""}"
vct_accuracy,Virology Capabilities Test,Safety Capability,accuracy (%),322.0,https://arxiv.org/abs/2504.16137,"{""harness"":""VCT capability harness"",""higher_is_better"":true,""judge"":""exact VCT accuracy"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Dual-use virology troubleshooting capability benchmark."",""range"":[0,100],""sampling"":""unknown"",""tools"":""multimodal input; no safeguards"",""version"":""VCT 322-item release""}"
biosecurity_vct_accuracy,Biosecurity VCT,Safety Capability,overall accuracy (%),,https://media.x.ai/v1/website/card-4p6-4cd2dc57.pdf,"{""harness"":""SpaceXAI internal harness"",""higher_is_better"":true,""judge"":""overall should-refuse and should-comply accuracy"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Internal task count is undisclosed."",""range"":[0,100],""sampling"":""unknown"",""tools"":""none; no safeguards"",""version"":""SpaceXAI internal Biosecurity VCT""}"
biousebench_severity5_refusal_rate,BioUseBench Severity-5,Safety,severity-5 refusal rate (%),,https://media.x.ai/v1/website/card-4p6-4cd2dc57.pdf,"{""harness"":""SpaceXAI internal harness"",""higher_is_better"":true,""judge"":""severity-5 refusal grader"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Internal high-elo dual-use biology split; task count is undisclosed."",""range"":[0,100],""sampling"":""unknown"",""tools"":""standard safeguards"",""version"":""SpaceXAI internal BioUseBench severity-5 split""}"
wmdp_cyber_accuracy,WMDP-Cyber,Safety Capability,accuracy (%),1987.0,https://wmdp.ai,"{""harness"":""WMDP official evaluation"",""higher_is_better"":true,""judge"":""multiple-choice exact accuracy"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Public WMDP-Cyber split contains 1,987 questions."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none; no safeguards"",""version"":""WMDP-Cyber public release""}"
xai_standard_jailbreak_compliance,SpaceXAI Standard Jailbreaks,Safety,compliance rate (%),,https://media.x.ai/v1/website/4p5-5184fdf9.pdf,"{""harness"":""SpaceXAI internal harness"",""higher_is_better"":false,""judge"":""should-refuse compliance grader"",""metric_type"":""rate_pct"",""multimodal_input"":false,""notes"":""Internal task count is undisclosed; lower is better."",""range"":[0,100],""sampling"":""unknown"",""tools"":""standard safeguards"",""version"":""SpaceXAI standard jailbreak suite""}"
strongreject_compliance,StrongREJECT Compliance,Safety,compliance rate (%),313.0,https://arxiv.org/abs/2402.10260,"{""harness"":""StrongREJECT evaluation"",""higher_is_better"":false,""judge"":""StrongREJECT compliance grader"",""metric_type"":""rate_pct"",""multimodal_input"":false,""notes"":""Public StrongREJECT suite contains 313 prompts; lower is better."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""standard safeguards"",""version"":""StrongREJECT 313-prompt suite""}"
xai_long_horizon_jailbreak_compliance,SpaceXAI Long-Horizon Jailbreaks,Safety,compliance rate (%),,https://media.x.ai/v1/website/card-4p6-4cd2dc57.pdf,"{""harness"":""SpaceXAI internal harness"",""higher_is_better"":false,""judge"":""long-horizon should-refuse compliance grader"",""metric_type"":""rate_pct"",""multimodal_input"":false,""notes"":""Includes multiturn, Crescendo, and malicious-use long-horizon attacks; count undisclosed."",""range"":[0,100],""sampling"":""unknown"",""tools"":""standard safeguards"",""version"":""SpaceXAI long-horizon jailbreak suite""}"
xai_general_refusal_compliance,SpaceXAI General Refusal Compliance,Safety,compliance rate (%),,https://media.x.ai/v1/website/4p5-5184fdf9.pdf,"{""harness"":""SpaceXAI internal harness"",""higher_is_better"":false,""judge"":""multilingual should-refuse compliance grader"",""metric_type"":""rate_pct"",""multimodal_input"":false,""notes"":""Internal multilingual task count is undisclosed; lower is better."",""range"":[0,100],""sampling"":""unknown"",""tools"":""standard safeguards"",""version"":""SpaceXAI multilingual general-refusal suite""}"
xai_child_safety_compliance,SpaceXAI Child Safety Compliance,Safety,compliance rate (%),,https://media.x.ai/v1/website/4p5-5184fdf9.pdf,"{""harness"":""SpaceXAI internal harness"",""higher_is_better"":false,""judge"":""CSAM and child-safety compliance grader"",""metric_type"":""rate_pct"",""multimodal_input"":false,""notes"":""Internal multi-turn suite; lower is better."",""range"":[0,100],""sampling"":""unknown"",""tools"":""standard safeguards"",""version"":""SpaceXAI CSAM and child-safety suite""}"
xai_bio_refusal_accuracy,SpaceXAI Bio Refusal,Safety,refusal accuracy (%),,https://media.x.ai/v1/website/4p5-5184fdf9.pdf,"{""harness"":""SpaceXAI internal harness"",""higher_is_better"":true,""judge"":""dangerous-query refusal accuracy"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Internal task count is undisclosed."",""range"":[0,100],""sampling"":""unknown"",""tools"":""full safeguards"",""version"":""SpaceXAI Autointent-Bio dangerous-query split""}"
xai_chem_refusal_accuracy,SpaceXAI Chem Refusal,Safety,refusal accuracy (%),,https://media.x.ai/v1/website/4p5-5184fdf9.pdf,"{""harness"":""SpaceXAI internal harness"",""higher_is_better"":true,""judge"":""dangerous-query refusal accuracy"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Internal task count is undisclosed."",""range"":[0,100],""sampling"":""unknown"",""tools"":""full safeguards"",""version"":""SpaceXAI Autointent-Chem dangerous-query split""}"
xai_rn_refusal_accuracy,SpaceXAI Radiological/Nuclear Refusal,Safety,refusal accuracy (%),,https://labs.scale.com/leaderboard/fortress,"{""harness"":""FORTRESS-RN evaluation"",""higher_is_better"":true,""judge"":""dangerous-query refusal accuracy"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The xAI card reports the R/N refusal row; exact evaluated count is undisclosed."",""range"":[0,100],""sampling"":""unknown"",""tools"":""full safeguards"",""version"":""FORTRESS-RN dangerous-query split""}"
xai_self_harm_compliance,SpaceXAI Self-Harm Compliance,Safety,compliance rate (%),,https://media.x.ai/v1/website/4p5-5184fdf9.pdf,"{""harness"":""SpaceXAI internal harness"",""higher_is_better"":false,""judge"":""self-harm assistance or failed-redirection compliance grader"",""metric_type"":""rate_pct"",""multimodal_input"":false,""notes"":""Internal task count is undisclosed; lower is better."",""range"":[0,100],""sampling"":""unknown"",""tools"":""standard safeguards"",""version"":""SpaceXAI self-harm suite""}"
xai_epistemic_bias_rate,SpaceXAI Epistemic Bias,Behavior,bias rate (%),,https://media.x.ai/v1/website/4p5-5184fdf9.pdf,"{""harness"":""SpaceXAI internal harness"",""higher_is_better"":false,""judge"":""opposing-framing bias grader"",""metric_type"":""rate_pct"",""multimodal_input"":false,""notes"":""Internal task count is undisclosed; lower is better."",""range"":[0,100],""sampling"":""unknown"",""tools"":""none"",""version"":""SpaceXAI epistemic-bias suite""}"
xai_mask_rectified_dishonesty,MASK-Rectified Dishonesty,Behavior,dishonesty rate (%),,https://www.mask-benchmark.ai/,"{""harness"":""SpaceXAI rectified harness"",""higher_is_better"":false,""judge"":""MASK-Rectified dishonesty grader"",""metric_type"":""rate_pct"",""multimodal_input"":false,""notes"":""Rectifies model-aware role-playing cases; exact evaluated count is undisclosed."",""range"":[0,100],""sampling"":""unknown"",""tools"":""none"",""version"":""SpaceXAI MASK-Rectified evaluation""}"
xai_sycophancy_rate,SpaceXAI Sycophancy,Behavior,sycophancy rate (%),,https://media.x.ai/v1/website/4p5-5184fdf9.pdf,"{""harness"":""SpaceXAI internal harness"",""higher_is_better"":false,""judge"":""accuracy-drop sycophancy grader"",""metric_type"":""rate_pct"",""multimodal_input"":false,""notes"":""Internal task count is undisclosed; lower is better."",""range"":[0,100],""sampling"":""unknown"",""tools"":""none"",""version"":""SpaceXAI internal sycophancy suite""}"
terminal_bench_3_0,Terminal-Bench 3.0,Agentic Coding,task success rate (%),74.0,https://www.tbench.ai/,"{""harness"":""Terminal-Bench 3.0 fixed harness"",""higher_is_better"":true,""judge"":""verified task-success evaluator"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Expanded successor to Terminal-Bench 2.1; public release contains 74 tasks and uses three trials by default. The xAI card does not disclose trial count, so its scores are noncanonical variants."",""range"":[0,100],""sampling"":""3 trials"",""tools"":""container terminal"",""version"":""Terminal-Bench 3.0 / FrontierBench 74-task release""}"
xai_inferenceeval_accuracy,SpaceXAI InferenceEval,AI R&D,accuracy (%),,https://media.x.ai/v1/website/card-4p6-4cd2dc57.pdf,"{""harness"":""SpaceXAI internal harness"",""higher_is_better"":true,""judge"":""hidden GPU unit score gated by integration probe"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Internal task count is undisclosed."",""range"":[0,100],""sampling"":""unknown"",""tools"":""Grok Build coding tools and GPU tests"",""version"":""SpaceXAI internal InferenceEval""}"
xai_kernelbenchinternal_accuracy_efficiency,SpaceXAI KernelBenchInternal,AI R&D,accuracy / efficiency score (%),,https://media.x.ai/v1/website/card-4p6-4cd2dc57.pdf,"{""harness"":""SpaceXAI internal harness"",""higher_is_better"":true,""judge"":""correctness and measured speedup composite"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Internal task count is undisclosed; the score combines correctness and speedup."",""range"":[0,100],""sampling"":""unknown"",""tools"":""sandboxed GPU coding tools"",""version"":""SpaceXAI internal KernelBench-inspired suite""}"
cve_bench_reward,CVE-Bench,Cybersecurity,reward (%),40.0,https://arxiv.org/abs/2503.17332,"{""harness"":""CVE-Bench sandbox"",""higher_is_better"":true,""judge"":""CVE-Bench exploit reward"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Public benchmark contains 40 CVE environments."",""range"":[0,100],""sampling"":""unknown"",""tools"":""unrestricted sandboxed web-application tools"",""version"":""CVE-Bench 40-environment release""}"
xai_securecodereview_reward,SpaceXAI SecureCodeReview,Cybersecurity,reward (%),,https://media.x.ai/v1/website/card-4p6-4cd2dc57.pdf,"{""harness"":""SpaceXAI internal harness"",""higher_is_better"":true,""judge"":""security-fix reward with regression checks"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Internal task count is undisclosed."",""range"":[0,100],""sampling"":""unknown"",""tools"":""fixed secure-code-review tools"",""version"":""SpaceXAI internal SecureCodeReview""}"
harvey_lab_vals_final_score,Harvey Legal Agent Benchmark (Vals),Professional,final score (%),,https://www.vals.ai/benchmarks/hlab,"{""harness"":""Vals Harvey LAB harness"",""higher_is_better"":true,""judge"":""expert all-pass final score"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Long-horizon legal work over client-matter files; count is undisclosed."",""range"":[0,100],""sampling"":""unknown"",""tools"":""Vals legal-agent file tools"",""version"":""Vals implementation of Harvey LAB""}"
workspace_bench_openclaw_100,Workspace Bench (100-task OpenClaw total),Agentic Office,total score (%),100.0,https://arxiv.org/abs/2605.03596,"{""harness"":""100-task OpenClaw setting"",""higher_is_better"":true,""judge"":""rubric-based evaluator"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official Seed2.1 Table 1 explicitly reports a 100-task OpenClaw setting. This source-defined subset is distinct from the full public Workspace-Bench release."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""OpenClaw workspace tools"",""version"":""Workspace Bench (100-task OpenClaw total)""}"
workspace_bench_openclaw_100_pass30,Workspace Bench (100-task OpenClaw Pass@30),Agentic Office,% tasks with rubric score >= 30,100.0,https://arxiv.org/abs/2605.03596,"{""harness"":""100-task OpenClaw setting"",""higher_is_better"":true,""judge"":""rubric-based evaluator"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Pass@threshold is the source's rubric-score threshold metric, not repeated sampling."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""OpenClaw workspace tools"",""version"":""Workspace Bench (100-task OpenClaw Pass@30)""}"
workspace_bench_openclaw_100_pass50,Workspace Bench (100-task OpenClaw Pass@50),Agentic Office,% tasks with rubric score >= 50,100.0,https://arxiv.org/abs/2605.03596,"{""harness"":""100-task OpenClaw setting"",""higher_is_better"":true,""judge"":""rubric-based evaluator"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Pass@threshold is the source's rubric-score threshold metric, not repeated sampling."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""OpenClaw workspace tools"",""version"":""Workspace Bench (100-task OpenClaw Pass@50)""}"
workspace_bench_openclaw_100_pass70,Workspace Bench (100-task OpenClaw Pass@70),Agentic Office,% tasks with rubric score >= 70,100.0,https://arxiv.org/abs/2605.03596,"{""harness"":""100-task OpenClaw setting"",""higher_is_better"":true,""judge"":""rubric-based evaluator"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Pass@threshold is the source's rubric-score threshold metric, not repeated sampling."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""OpenClaw workspace tools"",""version"":""Workspace Bench (100-task OpenClaw Pass@70)""}"
workspace_bench_openclaw_100_pass90,Workspace Bench (100-task OpenClaw Pass@90),Agentic Office,% tasks with rubric score >= 90,100.0,https://arxiv.org/abs/2605.03596,"{""harness"":""100-task OpenClaw setting"",""higher_is_better"":true,""judge"":""rubric-based evaluator"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Pass@threshold is the source's rubric-score threshold metric, not repeated sampling."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""OpenClaw workspace tools"",""version"":""Workspace Bench (100-task OpenClaw Pass@90)""}"
workspace_bench_openclaw_100_pass100,Workspace Bench (100-task OpenClaw Pass@100),Agentic Office,% tasks with rubric score >= 100,100.0,https://arxiv.org/abs/2605.03596,"{""harness"":""100-task OpenClaw setting"",""higher_is_better"":true,""judge"":""rubric-based evaluator"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Pass@threshold is the source's rubric-score threshold metric, not repeated sampling."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""OpenClaw workspace tools"",""version"":""Workspace Bench (100-task OpenClaw Pass@100)""}"
presentbench,PresentBench,Agentic Office,rubric score (%),238.0,https://arxiv.org/abs/2603.07244,"{""harness"":""official"",""higher_is_better"":true,""judge"":""fine-grained instance-specific rubrics"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Official PresentBench paper; exact scored task count pending metadata audit."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""slide-generation environment"",""version"":""PresentBench""}"
agent_startup_bench,Agent Startup Bench,Agentic Professional,expert-reviewed score (%),,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""official"",""higher_is_better"":true,""judge"":""expert review"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Seed-developed benchmark based on research and interviews with real AI-native startups; no public fixed count disclosed."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""research and professional-deliverable tools"",""version"":""Agent Startup Bench""}"
agents_last_exam_average_score,Agents' Last Exam (average overall score),Agentic,average overall score,,https://agents-last-exam.org/docs/ale/index.html,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""score"",""multimodal_input"":true,""notes"":""Distinct from the existing full-pass-rate metric."",""range"":null,""sampling"":""pass@1"",""tools"":""computer-use environment"",""version"":""Agents' Last Exam (average overall score)""}"
one_million_bench,OneMillion Bench,Agentic Professional,reported score (%),400.0,https://arxiv.org/abs/2603.07980,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official OneMillion-Bench paper; exact scored task count pending metadata audit."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""agentic professional-task environment"",""version"":""OneMillion Bench""}"
gdpval_seed_reported_score,GDPVal (Seed source-reported score),Agentic Professional,source-reported score (0-100),220.0,https://huggingface.co/datasets/openai/gdpval,"{""harness"":""official"",""higher_is_better"":true,""judge"":""source-specific evaluator; normalization undisclosed"",""metric_type"":""score"",""multimodal_input"":true,""notes"":""The Seed source reports 0-100 values without identifying AA Elo, OpenAI wins-or-ties, GDPVal-Diamond, or another published aggregation. A distinct metric prevents conflation."",""range"":null,""sampling"":""pass@1"",""tools"":""document/spreadsheet deliverable workflow"",""version"":""GDPVal (Seed source-reported score)""}"
xdailybench,xDailyBench,Agentic Daily Life,rubric score (%),,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""official"",""higher_is_better"":true,""judge"":""multi-dimensional rubric evaluator"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Seed-developed benchmark covering more than 30 vertical scenarios; no fixed public item count disclosed."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""benchmark-specified"",""version"":""xDailyBench""}"
doubao_multi_turn_bench,Doubao Multi-Turn Bench,Conversation,rubric score (%),,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""official"",""higher_is_better"":true,""judge"":""rubric-based evaluator"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Seed-developed benchmark filtered from real Doubao conversations; no fixed public item count disclosed."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""Doubao Multi-Turn Bench""}"
seedclawbench,SeedClawBench,Agentic,Agent-as-Judge score (%),100.0,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""official"",""higher_is_better"":true,""judge"":""Agent-as-Judge with read-only evidence tools"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""100 tasks constructed from more than 5,000 Ark tasks and more than 600 crowdsourced tasks; 89.69% human-machine rubric agreement on the pilot."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""OpenClaw-style tools and skills"",""version"":""SeedClawBench""}"
claw_eval_multimodal_pass3,Claw-Eval Multimodal (Pass^3),Multimodal Agentic,all-three-pass rate (%),303.0,https://raw.githubusercontent.com/claw-eval/claw-eval/9ac81fc3f18e9711bc0e0e3a96f0883887ae88b4/README.md,"{""harness"":""official Claw-Eval v1.1"",""higher_is_better"":true,""judge"":""full-trajectory benchmark grading"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Distinct 101-task multimodal subset with exactly 303 generated trajectories."",""range"":[0,100],""sampling"":""Pass^3: three successful trajectories required"",""tools"":""official multimodal Claw-Eval agent environment"",""version"":""Claw-Eval v1.1 multimodal Pass^3""}"
wildclaw_bench_60_openclaw,WildClawBench (60-task full OpenClaw suite),Multimodal Agentic,weighted overall score (%),60.0,https://github.com/InternLM/WildClawBench,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Full suite contains 35 text-oriented and 25 multimodal tasks; distinct from the existing 35-task text-only subset."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""multimodal agent harness"",""version"":""WildClawBench (60-task full OpenClaw suite)""}"
image2floorplan_avg_score,Image2FloorPlan (average graph-similarity score),Multimodal Spatial,average normalized graph-similarity score (%),,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""official"",""higher_is_better"":true,""judge"":""weighted room-connectivity graph similarity"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Seed-developed in-house benchmark; fixed item count undisclosed."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""high-resolution image perception"",""version"":""Image2FloorPlan (average graph-similarity score)""}"
mobileworld,MobileWorld,Computer Use,task success rate (%),201.0,https://arxiv.org/abs/2512.19432,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Seed excludes the MCP test split due to deployment issues."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""mobile GUI actions"",""version"":""MobileWorld""}"
creativework,CreativeWork,Computer Use,task success rate (%),,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Seed-developed benchmark; fixed item count undisclosed."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""GUI + MCP in Notion, Canva, and Figma"",""version"":""CreativeWork""}"
gameworld,GameWorld,Computer Use,task success rate (%),170.0,https://arxiv.org/abs/2604.07429,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Exact scored task count pending metadata audit."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""browser-game GUI actions"",""version"":""GameWorld""}"
program_bench_fully_resolved,ProgramBench (fully resolved),Coding,fully resolved tasks (%),200.0,https://programbench.com/,"{""harness"":""official ProgramBench sandbox"",""higher_is_better"":true,""judge"":""248,000+ behavioral tests"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Distinct from the macro-average behavioral-tests-passed metric."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""compiled executable and documentation"",""version"":""ProgramBench (fully resolved)""}"
program_bench_almost_resolved_95,ProgramBench (almost resolved >=95%),Coding,tasks with at least 95% pass rate (%),200.0,https://programbench.com/,"{""harness"":""official ProgramBench sandbox"",""higher_is_better"":true,""judge"":""248,000+ behavioral tests"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Distinct from fully resolved and macro-average metrics."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""compiled executable and documentation"",""version"":""ProgramBench (almost resolved >=95%)""}"
zerobench_sub_tools,ZeroBench subquestions (with tools),Multimodal Reasoning,accuracy (%),334.0,https://arxiv.org/abs/2502.09696,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Distinct from ZeroBench main with tools and the no-tool subquestion row."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""source tool-augmented setting"",""version"":""ZeroBench subquestions (with tools)""}"
swe_atlas,SWE-Atlas,Agentic Coding,task resolve rate (%),500.0,https://arxiv.org/abs/2605.08366,"{""harness"":""SWE-agent"",""higher_is_better"":true,""judge"":""test-based task resolution"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Patch-generation benchmark; distinct from SWE-Atlas-QnA."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""repository shell/editor"",""version"":""SWE-Atlas""}"
webbench,WebBench,Agentic Coding,live task completion rate (%),2454.0,https://github.com/Halluminate/WebBench,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Pin the filtered 2,454-task release, not the historical 5,750-task count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""browser workflows on 452 live sites"",""version"":""WebBench""}"
seedkernelbench_avg_speedup,SeedKernelBench (average speedup),Agentic Coding,average speedup ratio,,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""official"",""higher_is_better"":true,""judge"":""measured speedup"",""metric_type"":""score"",""multimodal_input"":false,""notes"":""Higher is better; source reports ratios such as 9.21x."",""range"":null,""sampling"":""pass@1"",""tools"":""GPU kernel implementation and benchmarking"",""version"":""SeedKernelBench (average speedup)""}"
trae_repo_env,Trae Repo Env,Agentic Coding,reported score (%),,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Seed/Trae internal benchmark; fixed item count undisclosed."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""Trae coding-agent environment"",""version"":""Trae Repo Env""}"
trae_artifacts,Trae Artifacts,Agentic Coding,reported score (%),,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Seed/Trae internal benchmark; fixed item count undisclosed."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""Trae coding-agent environment"",""version"":""Trae Artifacts""}"
trae_error_fix_python,Trae Error Fix Python,Agentic Coding,reported score (%),,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Seed/Trae internal benchmark; fixed item count undisclosed."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""Trae coding-agent environment"",""version"":""Trae Error Fix Python""}"
trae_error_fix_js,Trae Error Fix JavaScript,Agentic Coding,reported score (%),,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Seed/Trae internal benchmark; fixed item count undisclosed."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""Trae coding-agent environment"",""version"":""Trae Error Fix JavaScript""}"
trae_error_fix_java,Trae Error Fix Java,Agentic Coding,reported score (%),,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Seed/Trae internal benchmark; fixed item count undisclosed."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""Trae coding-agent environment"",""version"":""Trae Error Fix Java""}"
trae_error_fix_go,Trae Error Fix Go,Agentic Coding,reported score (%),,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Seed/Trae internal benchmark; fixed item count undisclosed."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""Trae coding-agent environment"",""version"":""Trae Error Fix Go""}"
trae_code_gen_python,Trae Code Gen Python,Agentic Coding,reported score (%),,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Seed/Trae internal benchmark; fixed item count undisclosed."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""Trae coding-agent environment"",""version"":""Trae Code Gen Python""}"
trae_code_gen_js,Trae Code Gen JavaScript,Agentic Coding,reported score (%),,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Seed/Trae internal benchmark; fixed item count undisclosed."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""Trae coding-agent environment"",""version"":""Trae Code Gen JavaScript""}"
frontiercs_overall_v1,FrontierCS v1 (overall),Agentic Research,mean continuous score (%),256.0,https://arxiv.org/abs/2512.15699,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Source-date v1 snapshot: 188 algorithmic + 68 research tasks. Overall is another metric view over the same 256 generations; do not add all three rows for cost aggregation."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""research and code-execution environment"",""version"":""FrontierCS v1 (overall)""}"
frontiercs_algorithmic_v1,FrontierCS v1 (algorithmic),Agentic Research,mean continuous score (%),188.0,https://arxiv.org/abs/2512.15699,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Source-date algorithmic track; excludes later FrontierCS 2.0 additions."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""research and code-execution environment"",""version"":""FrontierCS v1 (algorithmic)""}"
frontiercs_research_v1,FrontierCS v1 (research),Agentic Research,mean continuous score (%),68.0,https://arxiv.org/abs/2512.15699,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Source-date research/systems track with task-specific partial credit."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""research and code-execution environment"",""version"":""FrontierCS v1 (research)""}"
mathverse_vision_only,MathVerse (Vision-Only),Multimodal Math,accuracy (%),,https://github.com/ZrrSkywalker/MathVerse,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Vision-only MathVerse setting."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""MathVerse (Vision-Only)""}"
measurebench,MeasureBench,Multimodal Perception,average real-and-synthetic score (%),2442.0,https://arxiv.org/abs/2510.26865,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Source reports the average over real and synthetic subsets."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""MeasureBench""}"
worldbench,WorldBench,Multimodal Knowledge,accuracy (%),2000.0,https://arxiv.org/abs/2606.06538,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""WorldVQA paper benchmark family; exact WorldBench split count pending audit."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""WorldBench""}"
embspatial_bench,EmbSpatialBench,Multimodal Spatial,accuracy (%),3640.0,https://arxiv.org/abs/2406.05756,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Official paper/source resolution pending metadata audit."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""EmbSpatialBench""}"
kina,KINA,Knowledge,multiple-choice accuracy (%),899.0,https://arxiv.org/abs/2606.05104,"{""harness"":""official KINA/lighteval runner"",""higher_is_better"":true,""judge"":""letter extraction"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":"""",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""KINA""}"
msqa,Multicultural SimpleQA,Multilingual Knowledge,accuracy (%),1086.0,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Seed-developed internal benchmark across 11 major languages."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""Multicultural SimpleQA""}"
live_mathematician_bench_2026_06,LiveMathematicianBench (Seed2.1 June 2026 snapshot),Reasoning,reported score (%),,https://arxiv.org/abs/2604.01754,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Live benchmark; the exact Seed model-card snapshot count is not published. Do not silently substitute the later 608-item live release."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""LiveMathematicianBench (Seed2.1 June 2026 snapshot)""}"
fs_researcher,FS-Researcher,Agentic Research,reported score (%),,https://arxiv.org/abs/2602.01566,"{""harness"":""official"",""higher_is_better"":true,""judge"":""benchmark-specified"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Table 11 cites FS-Researcher, which is distinct from FrontierScience-Research in Tables 6-7."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""file-system-based research agent environment"",""version"":""FS-Researcher""}"
code_arena_frontend_elo,Code Arena: Frontend Elo,Coding,Arena Elo,,https://arena.ai/leaderboard/code,"{""harness"":""Arena Code leaderboard"",""higher_is_better"":true,""judge"":""human preference"",""metric_type"":""score"",""multimodal_input"":false,""notes"":""Historical 107,962-vote snapshot reproduced in the Seed release blog."",""range"":null,""sampling"":""dynamic leaderboard aggregation"",""tools"":""none"",""version"":""Code Arena: Frontend Elo""}"
seed21_claudecode_glm51_win_rate,Seed2.1 ClaudeCode vs GLM-5.1 (win rate),Human Preference / Internal,Seed model win rate (%),178.0,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""Seed ClaudeCode/Trae crowdsourced evaluation"",""higher_is_better"":true,""judge"":""anonymous developer preference or multidimensional human rating"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official Seed2.1 model-card internal evaluation."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""real-repository coding environment"",""version"":""Seed2.1 ClaudeCode vs GLM-5.1 (win rate)""}"
seed21_claudecode_glm51_net_win_rate,Seed2.1 ClaudeCode vs GLM-5.1 (net win rate),Human Preference / Internal,net win rate (percentage points),178.0,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""Seed ClaudeCode/Trae crowdsourced evaluation"",""higher_is_better"":true,""judge"":""anonymous developer preference or multidimensional human rating"",""metric_type"":""score"",""multimodal_input"":false,""notes"":""Official Seed2.1 model-card internal evaluation."",""range"":null,""sampling"":""pass@1"",""tools"":""real-repository coding environment"",""version"":""Seed2.1 ClaudeCode vs GLM-5.1 (net win rate)""}"
seed21_claudecode_opus46_win_rate,Seed2.1 ClaudeCode vs Claude Opus 4.6 (win rate),Human Preference / Internal,Seed model win rate (%),230.0,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""Seed ClaudeCode/Trae crowdsourced evaluation"",""higher_is_better"":true,""judge"":""anonymous developer preference or multidimensional human rating"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official Seed2.1 model-card internal evaluation."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""real-repository coding environment"",""version"":""Seed2.1 ClaudeCode vs Claude Opus 4.6 (win rate)""}"
seed21_claudecode_opus46_tie_rate,Seed2.1 ClaudeCode vs Claude Opus 4.6 (tie rate),Human Preference / Internal,tie rate (%),230.0,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""Seed ClaudeCode/Trae crowdsourced evaluation"",""higher_is_better"":false,""judge"":""anonymous developer preference or multidimensional human rating"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official Seed2.1 model-card internal evaluation."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""real-repository coding environment"",""version"":""Seed2.1 ClaudeCode vs Claude Opus 4.6 (tie rate)""}"
seed21_claudecode_opus46_loss_rate,Seed2.1 ClaudeCode vs Claude Opus 4.6 (loss rate),Human Preference / Internal,Seed model loss rate (%),230.0,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""Seed ClaudeCode/Trae crowdsourced evaluation"",""higher_is_better"":false,""judge"":""anonymous developer preference or multidimensional human rating"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official Seed2.1 model-card internal evaluation."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""real-repository coding environment"",""version"":""Seed2.1 ClaudeCode vs Claude Opus 4.6 (loss rate)""}"
seed21_claudecode_opus46_net_win_rate,Seed2.1 ClaudeCode vs Claude Opus 4.6 (net win rate),Human Preference / Internal,net win rate (percentage points),230.0,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""Seed ClaudeCode/Trae crowdsourced evaluation"",""higher_is_better"":true,""judge"":""anonymous developer preference or multidimensional human rating"",""metric_type"":""score"",""multimodal_input"":false,""notes"":""Official Seed2.1 model-card internal evaluation."",""range"":null,""sampling"":""pass@1"",""tools"":""real-repository coding environment"",""version"":""Seed2.1 ClaudeCode vs Claude Opus 4.6 (net win rate)""}"
seed21_trae_opus47_preference_win_rate,Seed2.1 Trae vs Claude Opus 4.7 (preference win rate),Human Preference / Internal,Seed model preference win rate (%),167.0,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""Seed ClaudeCode/Trae crowdsourced evaluation"",""higher_is_better"":true,""judge"":""anonymous developer preference or multidimensional human rating"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official Seed2.1 model-card internal evaluation."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""real-repository coding environment"",""version"":""Seed2.1 Trae vs Claude Opus 4.7 (preference win rate)""}"
seed21_trae_mean_rating,Seed2.1 Trae mean rating,Human Preference / Internal,mean rating across six dimensions,167.0,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""Seed ClaudeCode/Trae crowdsourced evaluation"",""higher_is_better"":true,""judge"":""anonymous developer preference or multidimensional human rating"",""metric_type"":""score"",""multimodal_input"":false,""notes"":""Official Seed2.1 model-card internal evaluation."",""range"":null,""sampling"":""pass@1"",""tools"":""real-repository coding environment"",""version"":""Seed2.1 Trae mean rating""}"
seed21_trae_fully_correct_rate,Seed2.1 Trae fully correct artifact rate,Human Preference / Internal,fully correct ready-to-use rate (%),167.0,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""Seed ClaudeCode/Trae crowdsourced evaluation"",""higher_is_better"":true,""judge"":""anonymous developer preference or multidimensional human rating"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official Seed2.1 model-card internal evaluation."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""real-repository coding environment"",""version"":""Seed2.1 Trae fully correct artifact rate""}"
seed21_trae_acceptable_delivery_rate,Seed2.1 Trae acceptable delivery rate,Human Preference / Internal,acceptable delivery rate (%),167.0,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""Seed ClaudeCode/Trae crowdsourced evaluation"",""higher_is_better"":true,""judge"":""anonymous developer preference or multidimensional human rating"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official Seed2.1 model-card internal evaluation."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""real-repository coding environment"",""version"":""Seed2.1 Trae acceptable delivery rate""}"
seed21_trae_severely_broken_rate,Seed2.1 Trae severely broken delivery rate,Human Preference / Internal,severely broken delivery rate (%),167.0,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""Seed ClaudeCode/Trae crowdsourced evaluation"",""higher_is_better"":false,""judge"":""anonymous developer preference or multidimensional human rating"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official Seed2.1 model-card internal evaluation."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""real-repository coding environment"",""version"":""Seed2.1 Trae severely broken delivery rate""}"
seed21_trae_delivery_completeness,Seed2.1 Trae delivery completeness,Human Preference / Internal,delivery completeness score,167.0,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""Seed ClaudeCode/Trae crowdsourced evaluation"",""higher_is_better"":true,""judge"":""anonymous developer preference or multidimensional human rating"",""metric_type"":""score"",""multimodal_input"":false,""notes"":""Official Seed2.1 model-card internal evaluation."",""range"":null,""sampling"":""pass@1"",""tools"":""real-repository coding environment"",""version"":""Seed2.1 Trae delivery completeness""}"
seed21_trae_fully_usable_rate,Seed2.1 Trae fully usable rate,Human Preference / Internal,fully usable rate (%),167.0,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""Seed ClaudeCode/Trae crowdsourced evaluation"",""higher_is_better"":true,""judge"":""anonymous developer preference or multidimensional human rating"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official Seed2.1 model-card internal evaluation."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""real-repository coding environment"",""version"":""Seed2.1 Trae fully usable rate""}"
seed21_trae_unusable_rate,Seed2.1 Trae unusable rate,Human Preference / Internal,unusable rate (%),167.0,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""Seed ClaudeCode/Trae crowdsourced evaluation"",""higher_is_better"":false,""judge"":""anonymous developer preference or multidimensional human rating"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official Seed2.1 model-card internal evaluation."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""real-repository coding environment"",""version"":""Seed2.1 Trae unusable rate""}"
seed21_trae_instruction_following,Seed2.1 Trae instruction-following score,Human Preference / Internal,instruction-following score,167.0,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""Seed ClaudeCode/Trae crowdsourced evaluation"",""higher_is_better"":true,""judge"":""anonymous developer preference or multidimensional human rating"",""metric_type"":""score"",""multimodal_input"":false,""notes"":""Official Seed2.1 model-card internal evaluation."",""range"":null,""sampling"":""pass@1"",""tools"":""real-repository coding environment"",""version"":""Seed2.1 Trae instruction-following score""}"
seed21_trae_boundary_adherence,Seed2.1 Trae boundary-adherence score,Human Preference / Internal,boundary-adherence score,167.0,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""Seed ClaudeCode/Trae crowdsourced evaluation"",""higher_is_better"":true,""judge"":""anonymous developer preference or multidimensional human rating"",""metric_type"":""score"",""multimodal_input"":false,""notes"":""Official Seed2.1 model-card internal evaluation."",""range"":null,""sampling"":""pass@1"",""tools"":""real-repository coding environment"",""version"":""Seed2.1 Trae boundary-adherence score""}"
seed21_dirtyfilter_validation_recall,Seed2.1 DirtyFilter validation recall,Data Cleaning Agent,validation recall (%),8037.0,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""300M-token autonomous rule-iteration budget"",""higher_is_better"":true,""judge"":""validation-set labels"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Figure 19; all models use the same harness and validation set."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""rule-only autonomous evaluator/optimizer loop"",""version"":""Seed2.1 DirtyFilter validation recall""}"
seed21_dirtyfilter_precision,Seed2.1 DirtyFilter precision,Data Cleaning Agent,precision (%),8037.0,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf,"{""harness"":""300M-token autonomous rule-iteration budget"",""higher_is_better"":true,""judge"":""validation-set labels"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Figure 19; all models use the same harness and validation set."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""rule-only autonomous evaluator/optimizer loop"",""version"":""Seed2.1 DirtyFilter precision""}"
atlas_mrcr_8needle_auc_1m,ATLAS MRCR 8-Needle AUC through 1M,Long Context,normalized length AUC of exact match (%),792.0,https://arxiv.org/html/2605.28079,"{""harness"":""ATLAS full eight-slice evaluation through 1M"",""higher_is_better"":true,""judge"":""deterministic benchmark-native evaluator"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Eight slices from 8K through 1M with 106/96/98/100/100/100/100/92 instances. Distinct from a pointwise or unweighted MRCR aggregate."",""range"":[0,100],""sampling"":""one model response per instance"",""tools"":""none"",""version"":""ATLAS full eight-slice MRCR 8-needle through 1M""}"
atlas_oolong_synth_auc_1m,ATLAS OOLong-Synth AUC through 1M,Long Context,normalized length AUC of answer-level score (%),800.0,https://arxiv.org/html/2605.28079,"{""harness"":""ATLAS full eight-slice evaluation through 1M"",""higher_is_better"":true,""judge"":""deterministic benchmark-native evaluator"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""100 instances at each of 8K, 16K, 32K, 64K, 128K, 256K, 512K and 1M; benchmark-native categorical/date/numeric/frequency answer scoring."",""range"":[0,100],""sampling"":""one model response per instance"",""tools"":""none"",""version"":""ATLAS full eight-slice OOLong-Synth through 1M""}"
atlas_graphwalks_extend_auc_1m,ATLAS GraphWalks Extend AUC through 1M,Long Context,normalized length AUC of node-set F1 (%),800.0,https://arxiv.org/html/2605.28079,"{""harness"":""ATLAS full eight-slice evaluation through 1M"",""higher_is_better"":true,""judge"":""deterministic benchmark-native evaluator"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official 128K and 1M data plus six ATLAS-generated slices using the official BFS/parent-node generation procedure."",""range"":[0,100],""sampling"":""one model response per instance"",""tools"":""none"",""version"":""ATLAS full eight-slice GraphWalks Extend through 1M""}"
atlas_loft_text_retrieval_extend_auc_1m,ATLAS LOFT Text Retrieval Extend AUC through 1M,Long Context,normalized length AUC of MRecall@K (%),800.0,https://arxiv.org/html/2605.28079,"{""harness"":""ATLAS full eight-slice evaluation through 1M"",""higher_is_better"":true,""judge"":""deterministic benchmark-native evaluator"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""LOFT retrieval plus HELMET-RAG evidence grounding. Remaining length slices are obtained by downsampling the 1M retrieval subset."",""range"":[0,100],""sampling"":""one model response per instance"",""tools"":""none"",""version"":""ATLAS full eight-slice LOFT Text Retrieval Extend through 1M""}"
atlas_helmet_icl_extend_auc_1m,ATLAS HELMET-ICL Extend AUC through 1M,Long Context,normalized length AUC of classification accuracy (%),800.0,https://arxiv.org/html/2605.28079,"{""harness"":""ATLAS full eight-slice evaluation through 1M"",""higher_is_better"":true,""judge"":""deterministic benchmark-native evaluator"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official HELMET construction scripts generate all ATLAS slices."",""range"":[0,100],""sampling"":""one model response per instance"",""tools"":""none"",""version"":""ATLAS full eight-slice HELMET-ICL Extend through 1M""}"
atlas_longcodeqa_auc_1m,ATLAS LongCodeQA AUC through 1M,Long Context,normalized length AUC of exact option accuracy (%),,https://arxiv.org/html/2605.28079,"{""harness"":""ATLAS six-slice evaluation from 32K through 1M"",""higher_is_better"":true,""judge"":""deterministic benchmark-native evaluator"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Table 15 reports the LongCodeQA subtask. ATLAS describes it as exact matching of the predicted option letter. Exact LongCodeQA counts are not separated from LongSWE in the published LongCodeBench composite counts."",""range"":[0,100],""sampling"":""one model response per instance"",""tools"":""none"",""version"":""ATLAS LongCodeQA six-slice evaluation from 32K through 1M""}"
atlas_amembench_acu_auc_1m,ATLAS AMemBench-ACU AUC through 1M,Long Context,normalized length AUC of QPEM (%),800.0,https://arxiv.org/html/2605.28079,"{""harness"":""ATLAS full eight-slice evaluation through 1M"",""higher_is_better"":true,""judge"":""deterministic benchmark-native evaluator"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""100 static transcript instances per slice; quasi-prefix exact match after normalization."",""range"":[0,100],""sampling"":""one model response per instance"",""tools"":""none"",""version"":""ATLAS full eight-slice AMemBench-ACU through 1M""}"
biology_instructions,Biology-Instructions,Science,source-reported aggregate score (%),,https://huggingface.co/datasets/SciReason/bio_instruction/tree/c536cf1a0727baba3c42c79972b8a64bb08c5488,"{""higher_is_better"":true,""judge"":""task-specific metrics"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Multi-omics task bundle across DNA, RNA, protein, and mixed tasks. Registry includes MCC, PCC, Spearman, Fmax, accuracy, AUC, R\u00b2, and mixed scores. One aggregate formula and one non-overlapping item count are not published in the score source; keep num_problems null."",""range"":[0,100],""sampling"":""pass@1/source-defined"",""tools"":""none"",""version"":""SciReason/bio_instruction at c536cf1a0727baba3c42c79972b8a64bb08c5488""}"
mol_instructions,Mol-Instructions,Science,source-reported aggregate score (%),,https://huggingface.co/datasets/SciReason/Mol-Instructions-test/tree/a581cc374ec90be8d808acf37d788f1f65d0395b,"{""higher_is_better"":true,""judge"":""task-specific exact/semantic metrics"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Heterogeneous molecular and protein instruction suite. Source table does not state a single aggregate formula or non-overlapping total; keep num_problems null."",""range"":[0,100],""sampling"":""pass@1/source-defined"",""tools"":""none"",""version"":""SciReason/Mol-Instructions-test at a581cc374ec90be8d808acf37d788f1f65d0395b""}"
moleculariq,MolecularIQ,Science,aggregate accuracy (%),5111.0,https://huggingface.co/datasets/ml-jku/moleculariq-v0.0/tree/aa3d7c6c2a67c20977f3fefa5169e65829166450,"{""higher_is_better"":true,""judge"":""task-specific rule-based scoring"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Official dataset card reports a 5,111-example test split. Additional task-specific splits overlap/derive from the benchmark and are not added to the canonical count."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""moleculariq-v0.0 test split""}"
scireasoner,SciReasoner,Science,source-reported aggregate score (%),,https://arxiv.org/abs/2509.21320v3,"{""higher_is_better"":true,""judge"":""task-specific scientific reasoning metrics"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Broad suite spanning chemistry, biology, materials, sequence, prediction, classification, generation, and extraction. Examined official sources do not publish one aggregate item count/formula; keep num_problems null."",""range"":[0,100],""sampling"":""source-defined"",""tools"":""none"",""version"":""SciReasoner evaluation suite in arXiv:2509.21320v3""}"
hle_multimodal,HLE Multimodal,Multimodal,% correct,342.0,https://huggingface.co/datasets/cais/hle/tree/5a81a4c7271a2a2a312b9a690f0c2fde837e4c29,"{""higher_is_better"":true,""judge"":""official HLE answer evaluation"",""metric_type"":""pct"",""multimodal_input"":true,""notes"":""Count is 2,500 finalized HLE minus the campaign-established 2,158 text-only rows = 342. Must not reuse hle_vl, which is a different search-enabled setting."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""none"",""version"":""Multimodal-only subset of finalized 2,500-question HLE""}"
lmarena_search_elo,LMArena Search Arena Elo,Search Agent,Arena score,,https://arena.ai/leaderboard/search,"{""higher_is_better"":true,""judge"":""human pairwise preference votes"",""metric_type"":""elo"",""multimodal_input"":false,""notes"":""Dated live-leaderboard score, distinct from text-only chatbot_arena_elo. No fixed static item set."",""range"":null,""sampling"":""live pairwise battles"",""style_control"":""off"",""tools"":""search/grounding environment varies by model"",""version"":""LMArena Search live Arena score""}"
aa_lcr_mistral_custom_gpt_4_1_mini_middle_out,AA Long Context Reasoning (Mistral custom),Long Context,% correct,100.0,https://huggingface.co/mistralai/Mistral-Small-4-119B-2603/blob/a11f36bebf709121056b1dbcc943d1c6afbe494d/README.md,"{""context_handling"":""middle-out for models with shorter context lengths"",""harness"":""Mistral custom AA-LCR implementation"",""higher_is_better"":true,""judge"":""gpt-4.1-mini-2025-04-14 equality judge"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Distinct from campaign aa_lcr: Mistral changes the judge and adds middle-out context handling. The underlying AA-LCR set has 100 questions; Mistral does not disclose repeats or total generations."",""prompt_style"":""AA-LCR 100 hard open-answer questions over approximately 100k-token inputs; exact Mistral prompt undisclosed"",""range"":[0,100],""sampling"":""pass@1; repeat count not disclosed by Mistral"",""temperature"":""undisclosed"",""tools"":""none"",""version"":""Mistral Small 4 custom implementation of AA-LCR""}"
aa_omniscience_index,AA Omniscience Index,Factuality,index (-100 to 100),60000.0,https://artificialanalysis.ai/evaluations/omniscience,"{""higher_is_better"":true,""metric_type"":""index"",""multimodal_input"":false,""notes"":""Rewards correct answers and penalizes hallucinations."",""range"":[-100,100],""version"":""AA-Omniscience 6,000 questions x 10 runs""}"
aa_omniscience_public_index,AA Omniscience Public Index,Factuality,index (-100 to 100),,https://www.liquid.ai/blog/lfm2-5-2-6b,"{""higher_is_better"":true,""metric_type"":""index"",""multimodal_input"":false,""notes"":""Public subset/index reported by Liquid AI; item count is not stated."",""range"":[-100,100],""version"":""AA-Omniscience-Public""}"
aa_omniscience_public_accuracy,AA Omniscience Public Accuracy,Factuality,% correct,,https://huggingface.co/LiquidAI/LFM2.5-2.6B,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Public-subset accuracy reported by Liquid AI."",""range"":[0,100],""version"":""AA-Omniscience-Public""}"
aa_omniscience_public_non_hallucination,AA Omniscience Public Non-Hallucination,Hallucination,% non-hallucination,,https://huggingface.co/LiquidAI/LFM2.5-2.6B,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Public-subset non-hallucination rate reported by Liquid AI."",""range"":[0,100],""version"":""AA-Omniscience-Public""}"
ifstruct_v1,IFStruct v1.0,Instruction Following,% binary structural compliance,2000.0,https://huggingface.co/datasets/LiquidAI/ifstruct-v1.0,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""No constrained decoding; JSON/YAML structure and schema only."",""range"":[0,100],""version"":""IFStruct v1.0 frozen public test set""}"
tool_sandbox,ToolSandbox,Tool Use,% aggregate milestone similarity,,https://github.com/apple/ToolSandbox,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Stateful conversational tool use with milestone/minefield scoring."",""range"":[0,100],""version"":""ToolSandbox arXiv:2408.04682v2""}"
claw_eval_en_average,Claw-Eval English Average,Agentic,Pass^3 average (%),179.0,https://huggingface.co/datasets/claw-eval/Claw-Eval,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""101 general, 72 multimodal, and 6 multi-turn English tasks; N=3."",""range"":[0,100],""version"":""Claw-Eval public English tasks""}"
browsecomp_plus_openclaw,BrowseComp-Plus (OpenClaw),Agentic Search,% correct,830.0,https://huggingface.co/datasets/Tevatron/browsecomp-plus,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""OpenClaw agent over the fixed approximately 100K-document corpus."",""range"":[0,100],""version"":""BrowseComp-Plus fixed-corpus test set""}"
air_bench_2024,AIR-Bench 2024,Safety,safe-engagement score (%),,https://arxiv.org/abs/2407.17436,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Category-specific LLM judges score policy-grounded safe engagement across regulatory and policy-derived harms."",""range"":[0,100],""version"":""AIR-Bench 2024 aggregate""}"
cyberseceval4_instruct,CyberSecEval 4 Instruct,Safety,secure-code score (%),,https://github.com/meta-llama/PurpleLlama,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Static-analysis evaluation of responses to coding requests designed to elicit known insecure patterns."",""range"":[0,100],""version"":""CyberSecEval 4 insecure-code-generation Instruct""}"
cyberseceval4_autocomplete,CyberSecEval 4 Autocomplete,Safety,secure-code score (%),,https://github.com/meta-llama/PurpleLlama,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Static-analysis evaluation of code completions where context leads up to a known insecure pattern."",""range"":[0,100],""version"":""CyberSecEval 4 insecure-code-generation Autocomplete""}"
longfact_claim_precision,LongFact Claim Precision,Factuality,claim-level precision (%),,https://arxiv.org/abs/2403.18802,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Uses the LongFact prompt set and the simplified claim extraction plus LLM-judge protocol described by Microsoft."",""range"":[0,100],""version"":""LongFact prompts with simplified claim extraction""}"
corpusqa_gpt54_judge,CorpusQA with GPT-5.4 Judge,Long Context,AI-judge score (%),1316.0,https://arxiv.org/abs/2601.14952,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""1,316 multi-document free-form QA instances. Microsoft replaces the default DeepSeek-V3 judge with GPT-5.4 high."",""range"":[0,100],""version"":""CorpusQA with Microsoft GPT-5.4-high judge""}"
amo_bench,AMO Bench,Math,accuracy (%),,https://microsoft.ai/pdf/MAI-Code-1-Flash-Model-Card.PDF,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Olympiad-math benchmark; public task count is not stated."",""range"":[0,100],""version"":""MAI-Code-1-Flash model-card release evaluation""}"
advancedif_rubric_level,AdvancedIF Rubric-Level Score,Instruction Following,rubric-level score (%),1645.0,https://arxiv.org/abs/2511.10507,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Microsoft explicitly reports rubric-level scores, distinct from the all-rubrics-pass benchmark primary."",""range"":[0,100],""version"":""AdvancedIF public test split; rubric-level aggregation""}"
advancedif_average,AdvancedIF Average,Instruction Following,source-reported average score (%),1645.0,https://arxiv.org/abs/2511.10507,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""MAI-Code-1-Flash chart labels this metric as an average, not the all-rubrics-pass primary."",""range"":[0,100],""version"":""AdvancedIF public test split; source-reported average""}"
ifbench_single_multiturn_average,IFBench Single/Multi-Turn Average,Instruction Following,single/multi-turn average score (%),1687.0,https://arxiv.org/abs/2507.02833,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""300 single-turn plus 1,387 multi-turn examples. MAI-Code-1-Flash reports the average of the two aggregate evaluation scores."",""range"":[0,100],""version"":""IFBench single-turn and multi-turn average""}"
truthfulqa_mc,TruthfulQA Multiple Choice,Factuality,multiple-choice accuracy (%),817.0,https://github.com/sylinrl/TruthfulQA,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Distinct from the generation benchmark already registered as truthfulqa."",""range"":[0,100],""version"":""TruthfulQA recommended multiple-choice setting""}"
longbench_v2_256k,LongBench-V2 256K Subset,Long Context,multiple-choice accuracy (%),408.0,https://huggingface.co/datasets/THUDM/LongBench-v2,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Microsoft follows the official setup but limits input context to 256K, leaving 408 unique questions."",""range"":[0,100],""version"":""LongBench-V2 questions fitting within 256K input context""}"
graphwalks_bfs_bucket_0_4k,GraphWalks BFS bucket 0–4K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact bucket identity from the Xiaomi embedded JavaScript data and the immutable official GraphWalks README after the 2026-02-27 parents-ground-truth and BFS-prompt fixes. The source reports cell-specific n values, which are preserved in each observation setting rather than inventing a benchmark-wide count."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; BFS exact context bucket 0\u20134K""}"
graphwalks_bfs_bucket_4k_8k,GraphWalks BFS bucket 4K–8K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact bucket identity from the Xiaomi embedded JavaScript data and the immutable official GraphWalks README after the 2026-02-27 parents-ground-truth and BFS-prompt fixes. The source reports cell-specific n values, which are preserved in each observation setting rather than inventing a benchmark-wide count."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; BFS exact context bucket 4K\u20138K""}"
graphwalks_bfs_bucket_8k_16k,GraphWalks BFS bucket 8K–16K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact bucket identity from the Xiaomi embedded JavaScript data and the immutable official GraphWalks README after the 2026-02-27 parents-ground-truth and BFS-prompt fixes. The source reports cell-specific n values, which are preserved in each observation setting rather than inventing a benchmark-wide count."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; BFS exact context bucket 8K\u201316K""}"
graphwalks_bfs_bucket_16k_32k,GraphWalks BFS bucket 16K–32K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact bucket identity from the Xiaomi embedded JavaScript data and the immutable official GraphWalks README after the 2026-02-27 parents-ground-truth and BFS-prompt fixes. The source reports cell-specific n values, which are preserved in each observation setting rather than inventing a benchmark-wide count."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; BFS exact context bucket 16K\u201332K""}"
graphwalks_bfs_bucket_32k_64k,GraphWalks BFS bucket 32K–64K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact bucket identity from the Xiaomi embedded JavaScript data and the immutable official GraphWalks README after the 2026-02-27 parents-ground-truth and BFS-prompt fixes. The source reports cell-specific n values, which are preserved in each observation setting rather than inventing a benchmark-wide count."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; BFS exact context bucket 32K\u201364K""}"
graphwalks_bfs_bucket_64k_128k,GraphWalks BFS bucket 64K–128K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact bucket identity from the Xiaomi embedded JavaScript data and the immutable official GraphWalks README after the 2026-02-27 parents-ground-truth and BFS-prompt fixes. The source reports cell-specific n values, which are preserved in each observation setting rather than inventing a benchmark-wide count."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; BFS exact context bucket 64K\u2013128K""}"
graphwalks_bfs_bucket_128k_256k,GraphWalks BFS bucket 128K–256K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact bucket identity from the Xiaomi embedded JavaScript data and the immutable official GraphWalks README after the 2026-02-27 parents-ground-truth and BFS-prompt fixes. The source reports cell-specific n values, which are preserved in each observation setting rather than inventing a benchmark-wide count."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; BFS exact context bucket 128K\u2013256K""}"
graphwalks_bfs_bucket_256k_512k,GraphWalks BFS bucket 256K–512K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact bucket identity from the Xiaomi embedded JavaScript data and the immutable official GraphWalks README after the 2026-02-27 parents-ground-truth and BFS-prompt fixes. The source reports cell-specific n values, which are preserved in each observation setting rather than inventing a benchmark-wide count."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; BFS exact context bucket 256K\u2013512K""}"
graphwalks_parents_bucket_0_4k,GraphWalks Parents bucket 0–4K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact bucket identity from the Xiaomi embedded JavaScript data and the immutable official GraphWalks README after the 2026-02-27 parents-ground-truth and BFS-prompt fixes. The source reports cell-specific n values, which are preserved in each observation setting rather than inventing a benchmark-wide count."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; Parents exact context bucket 0\u20134K""}"
graphwalks_parents_bucket_4k_8k,GraphWalks Parents bucket 4K–8K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact bucket identity from the Xiaomi embedded JavaScript data and the immutable official GraphWalks README after the 2026-02-27 parents-ground-truth and BFS-prompt fixes. The source reports cell-specific n values, which are preserved in each observation setting rather than inventing a benchmark-wide count."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; Parents exact context bucket 4K\u20138K""}"
graphwalks_parents_bucket_8k_16k,GraphWalks Parents bucket 8K–16K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact bucket identity from the Xiaomi embedded JavaScript data and the immutable official GraphWalks README after the 2026-02-27 parents-ground-truth and BFS-prompt fixes. The source reports cell-specific n values, which are preserved in each observation setting rather than inventing a benchmark-wide count."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; Parents exact context bucket 8K\u201316K""}"
graphwalks_parents_bucket_16k_32k,GraphWalks Parents bucket 16K–32K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact bucket identity from the Xiaomi embedded JavaScript data and the immutable official GraphWalks README after the 2026-02-27 parents-ground-truth and BFS-prompt fixes. The source reports cell-specific n values, which are preserved in each observation setting rather than inventing a benchmark-wide count."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; Parents exact context bucket 16K\u201332K""}"
graphwalks_parents_bucket_32k_64k,GraphWalks Parents bucket 32K–64K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact bucket identity from the Xiaomi embedded JavaScript data and the immutable official GraphWalks README after the 2026-02-27 parents-ground-truth and BFS-prompt fixes. The source reports cell-specific n values, which are preserved in each observation setting rather than inventing a benchmark-wide count."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; Parents exact context bucket 32K\u201364K""}"
graphwalks_parents_bucket_64k_128k,GraphWalks Parents bucket 64K–128K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact bucket identity from the Xiaomi embedded JavaScript data and the immutable official GraphWalks README after the 2026-02-27 parents-ground-truth and BFS-prompt fixes. The source reports cell-specific n values, which are preserved in each observation setting rather than inventing a benchmark-wide count."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; Parents exact context bucket 64K\u2013128K""}"
graphwalks_parents_bucket_128k_256k,GraphWalks Parents bucket 128K–256K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact bucket identity from the Xiaomi embedded JavaScript data and the immutable official GraphWalks README after the 2026-02-27 parents-ground-truth and BFS-prompt fixes. The source reports cell-specific n values, which are preserved in each observation setting rather than inventing a benchmark-wide count."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; Parents exact context bucket 128K\u2013256K""}"
graphwalks_parents_bucket_256k_512k,GraphWalks Parents bucket 256K–512K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact bucket identity from the Xiaomi embedded JavaScript data and the immutable official GraphWalks README after the 2026-02-27 parents-ground-truth and BFS-prompt fixes. The source reports cell-specific n values, which are preserved in each observation setting rather than inventing a benchmark-wide count."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; Parents exact context bucket 256K\u2013512K""}"
graphwalks_bfs_32k,GraphWalks BFS at 32K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact-length point from the locked Xiaomi line chart. The benchmark definition is the immutable official GraphWalks README after the 2026-02-27 fixes. The chart prints F1 on [0,1]; campaign scores are the same values expressed as percentages."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; BFS exact 32K point""}"
graphwalks_bfs_64k,GraphWalks BFS at 64K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact-length point from the locked Xiaomi line chart. The benchmark definition is the immutable official GraphWalks README after the 2026-02-27 fixes. The chart prints F1 on [0,1]; campaign scores are the same values expressed as percentages."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; BFS exact 64K point""}"
graphwalks_bfs_128k,GraphWalks BFS at 128K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact-length point from the locked Xiaomi line chart. The benchmark definition is the immutable official GraphWalks README after the 2026-02-27 fixes. The chart prints F1 on [0,1]; campaign scores are the same values expressed as percentages."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; BFS exact 128K point""}"
graphwalks_bfs_512k,GraphWalks BFS at 512K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact-length point from the locked Xiaomi line chart. The benchmark definition is the immutable official GraphWalks README after the 2026-02-27 fixes. The chart prints F1 on [0,1]; campaign scores are the same values expressed as percentages."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; BFS exact 512K point""}"
graphwalks_parents_32k,GraphWalks Parents at 32K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact-length point from the locked Xiaomi line chart. The benchmark definition is the immutable official GraphWalks README after the 2026-02-27 fixes. The chart prints F1 on [0,1]; campaign scores are the same values expressed as percentages."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; Parents exact 32K point""}"
graphwalks_parents_64k,GraphWalks Parents at 64K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact-length point from the locked Xiaomi line chart. The benchmark definition is the immutable official GraphWalks README after the 2026-02-27 fixes. The chart prints F1 on [0,1]; campaign scores are the same values expressed as percentages."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; Parents exact 64K point""}"
graphwalks_parents_128k,GraphWalks Parents at 128K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact-length point from the locked Xiaomi line chart. The benchmark definition is the immutable official GraphWalks README after the 2026-02-27 fixes. The chart prints F1 on [0,1]; campaign scores are the same values expressed as percentages."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; Parents exact 128K point""}"
graphwalks_parents_256k,GraphWalks Parents at 256K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact-length point from the locked Xiaomi line chart. The benchmark definition is the immutable official GraphWalks README after the 2026-02-27 fixes. The chart prints F1 on [0,1]; campaign scores are the same values expressed as percentages."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; Parents exact 256K point""}"
graphwalks_parents_512k,GraphWalks Parents at 512K,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact-length point from the locked Xiaomi line chart. The benchmark definition is the immutable official GraphWalks README after the 2026-02-27 fixes. The chart prints F1 on [0,1]; campaign scores are the same values expressed as percentages."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; Parents exact 512K point""}"
graphwalks_parents_1m,GraphWalks Parents at 1M,Long Context,% F1,,https://huggingface.co/datasets/openai/graphwalks/resolve/f338bb265735a56a79f4b0f5def722c9c3268ead/README.md,"{""higher_is_better"":true,""judge"":""deterministic set-overlap F1 against answer node list"",""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exact-length point from the locked Xiaomi line chart. The benchmark definition is the immutable official GraphWalks README after the 2026-02-27 fixes. The chart prints F1 on [0,1]; campaign scores are the same values expressed as percentages."",""range"":[0,100],""sampling"":""one response per graph prompt"",""tools"":""none"",""version"":""GraphWalks post-2026-02-27 ground-truth/prompt fix; Parents exact 1M point""}"
livecodebench_v6_2408_2505,LiveCodeBench v6 (2024-08 to 2025-05),Coding,pass@1 / mean pass@1 (%),454.0,https://raw.githubusercontent.com/LiveCodeBench/LiveCodeBench/28fef95ea8c9f7a547c8329f2cd3d32b92c1fa24/README.md,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The pinned official LiveCodeBench repository defines release versions and date-window filtering. The InclusionAI report is score evidence for this exact 454-problem 2024-08 to 2025-05 window, distinct from the 1,055-problem release_v6."",""range"":[0,100],""version"":""LiveCodeBench v6, 454 problems, 2024-08 through 2025-05""}"
claw_eval_general_pass3,Claw-Eval General (Pass^3),Agentic,all-three-pass rate (%),483.0,https://raw.githubusercontent.com/claw-eval/claw-eval/9ac81fc3f18e9711bc0e0e3a96f0883887ae88b4/README.md,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""Exactly 161 general tasks x 3 trajectories. Distinct from the 199-task non-multimodal general+multi-turn aggregate and from the 101-task multimodal split."",""range"":[0,100],""version"":""Claw-Eval v1.1 general split, 161 tasks, three trials""}"
pinchbench_123,PinchBench 123-task manifest at commit 27afe091b6ae04ec4b6aa9f5459bd280da0fd61d (2026-04-24),Agentic,mean task score (%),123.0,https://raw.githubusercontent.com/pinchbench/skill/27afe091b6ae04ec4b6aa9f5459bd280da0fd61d/manifest.yaml,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":false,""notes"":""The immutable manifest has 123 unique task IDs. Each task is scored on [0,1], including partial rubric credit; a run score is the arithmetic mean across tasks. Best Score and average across runs are distinct aggregations. No release tag is inferred. This is distinct from the campaign's 53-task identity."",""range"":[0,100],""version"":""PinchBench 123-task manifest at commit 27afe091b6ae04ec4b6aa9f5459bd280da0fd61d (2026-04-24)""}"
vstar,V* Bench,Vision,multiple-choice accuracy (%),191.0,https://huggingface.co/datasets/craigwu/vstar_bench/tree/d9ae62c903da0c98336e85c5ee89cd863b04b4da,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""The pinned official test_questions.jsonl has 191 items. StepFun's reported V* cells use a Python visual tool and are therefore non-default score settings, not a distinct benchmark."",""range"":[0,100],""tools"":""none; score-level settings may add a Python visual-search tool"",""version"":""V* Bench official 191-question test_questions release""}"
hr_bench_4k,HR-Bench 4K,Vision,accuracy (%),800.0,https://huggingface.co/datasets/DreamMr/HR-Bench/tree/83b9013d6293b85dc507e87199ca52517536939c,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""The benchmark has 200 logical questions: 100 FSP and 100 FCP. The official 4K split contains 800 evaluated rows after four cyclic answer-option rotations, so num_problems records 800 physical model generations."",""range"":[0,100],""tools"":""none; score-level settings may add a Python visual tool"",""version"":""HR-Bench 4K official option-rotation release""}"
hr_bench_8k,HR-Bench 8K,Vision,accuracy (%),800.0,https://huggingface.co/datasets/DreamMr/HR-Bench/tree/83b9013d6293b85dc507e87199ca52517536939c,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""The benchmark has 200 logical questions: 100 FSP and 100 FCP. The official 8K split contains 800 evaluated rows after four cyclic answer-option rotations, so num_problems records 800 physical model generations."",""range"":[0,100],""tools"":""none; score-level settings may add a Python visual tool"",""version"":""HR-Bench 8K official option-rotation release""}"
android_daily,AndroidDaily,GUI Agent,pass@1 task success rate (%),350.0,https://arxiv.org/html/2605.27761v1,"{""higher_is_better"":true,""metric_type"":""pct"",""multimodal_input"":true,""notes"":""The official paper defines 350 tasks over 94 applications and reports pass@1 success without multi-seed variance. Each rollout may run for up to 40 minutes. StepFun's launch chart uses its Step-specific phone-use stack; that harness remains score-level provenance."",""range"":[0,100],""tools"":""physical Android phone-use action environment"",""version"":""AndroidDaily v1: 350 tasks across 94 closed-source applications""}"
osworld_2_0_2026_06_24_partial,OSWorld 2.0 (2026-06-24 partial score),Agentic,weighted checkpoint partial score (%),108.0,https://raw.githubusercontent.com/xlang-ai/OSWorld-V2/8b6b59660b59832a42a345db8f86fa9f98c37573/benchmark_releases/osworld-v2-2026.06.24.json,"{""higher_is_better"":true,""metric_type"":""weighted_checkpoint_score_pct"",""notes"":""Distinct official 108-task release manifest from 2026-06-24, before the 2026-08-08 patch. The metric is the weighted checkpoint partial score."",""range"":[0,100],""version"":""osworld-v2-2026.06.24, before 2026-08-08 patch""}"
labbench2,LABBench2,Science,accuracy (%),1912.0,https://raw.githubusercontent.com/EdisonScientific/labbench2/c028ecdcf144b55ffcd92b68be45081df5628c20/README.md,"{""dataset_split"":""full 1,912-task suite"",""harness"":""official LABBench2 evaluation harness"",""higher_is_better"":true,""judge"":""task-specific exact/programmatic evaluator"",""metric_type"":""accuracy_pct"",""multimodal_input"":true,""notes"":""Paper Table 1 totals 1,912 evaluation units across literature, databases, protocols, sequence, cloning, figure, and table tasks."",""range"":[0,100],""sampling"":""pass@1"",""tools"":""benchmark/model-specific"",""version"":""LABBench2 full 1,912-task release""}"