amazon/ShoppingReasoningBench

Shopping Reasoning Bench

Conversational shopping assistants now serve hundreds of millions of customers, yet no existing benchmark jointly evaluates the open-ended multi-turn reasoning, domain expertise, and criterion-level quality that real shopping conversations demand. We introduce the Shopping Reasoning Bench, an expert-authored benchmark of 525 missions (232 single-turn, 293 multi-turn) with 10,863 importance-weighted binary rubrics authored by retail domain experts. These criteria are organized under a taxonomy of five reasoning categories and fifteen subcategories covering diverse demands such as preference refinement, trade-off analysis, and compatibility assessment.

Shopping reasoning is unique among language model applications. Unlike factual question answering or verifiable code generation, it requires balancing subjective preferences, budget constraints, and cross-product trade-offs across multi-turn dialogue, capabilities absent from previous e-commerce and general-purpose benchmarks.

• 📄 Paper: Shopping Reasoning Bench: An Expert-Authored Benchmark for Multi-Turn Conversational Shopping Assistants
• 🏷️ License: CC-BY-NC-4.0 (research / non-commercial use)

Dataset Summary

• Expert-authored multi-turn shopping dataset. 232 single-turn queries and 293 multi-turn missions (1,764 turns) authored by retail domain experts across five product families.
• Importance-weighted atomic rubric framework. 10,863 binary criteria (85.0% required) that decompose expert shopping reasoning into independently verifiable pass/fail checks.
• First taxonomy of pre-purchase shopping reasoning. Five categories and fifteen subcategories grounded in expert-annotated turns (Product Recommendation, Shopping Guidance, Product Comparison, Product Inquiry, Conversational Navigation).
• Validated LLM-as-judge. A single LLM judge applies uniform decision criteria across the benchmark; its reliability is validated against expert annotations. The judge prompt is included.

Dataset Structure

The benchmark is released as four configurations, each a single test split. The full_* configs are the complete benchmark of 525 missions and 10,863 rubrics. The hard_* configs are a focused Shopping Reasoning Bench-Hard subset of the 108 missions where the nine-model average weighted pass rate falls below 60%, representing the missions that current models collectively struggle with; it is designed for tracking progress on the most demanding shopping reasoning problems.

Config	Missions	Turns	Rubrics
full_single_turn	232	232	821
full_multi_turn	293	1,764	10,042
hard_single_turn	69	69	252
hard_multi_turn	39	235	1,411

Each line is one mission: top-level metadata (mission_id, mission_type, product_family, …) plus a list of turns. Each turn carries the customer messages and a list of rubrics, where each rubric has a text, an importance (required / optional), and taxonomy tags. See the paper (Appendix G) for the full field reference.

Usage

from datasets import load_dataset

ds = load_dataset("amazon/ShoppingReasoningBench", "full_multi_turn", split="test")

mission = ds[0]
print(mission["mission_name"])
for turn in mission["turns"]:
    print(turn["messages"][-1]["content"])
    for rubric in turn["rubrics"]:
        print(f"  [{rubric['importance']}] {rubric['text']}")

Configs: full_single_turn, full_multi_turn, hard_single_turn, hard_multi_turn.

Evaluation

Shopping Reasoning Bench aggregates atomic rubric judgments into per-turn, per-mission, and dataset-level scores via importance-weighted pass rates. Each rubric carries an importance weight (w_i = 5 for required rubrics, w_i = 1 for optional), and the weighted pass rate for a response is the sum of weights of passed rubrics over the sum of all weights. Scores aggregate hierarchically: the per-mission score is the mean of its per-turn scores, and the dataset-level score is the mean of per-mission scores.

Each rubric is scored by a single LLM judge that produces a binary pass/fail decision with a brief rationale; the paper uses Claude Sonnet 4.5 as the judge at temperature 0. The judge prompt is released as judge_prompt.txt; its <<current_conversation>>, <<rubric_text>>, and <<conversation_history>> placeholders are filled at evaluation time with the current turn, rubric text, and (for multi-turn) conversation history through the preceding turn.

Citation

@article{fan2026shopping,
  title={Shopping Reasoning Bench: An Expert-Authored Benchmark for Multi-Turn Conversational Shopping Assistants},
  author={Fan, Shuxian and Min, Seonwoo and Hu, Youna and Xia, Botao and Unnikrishnan, Jayakrishnan and Musselmann, Rowan and Gao, Yifan and Yin, Qingyu and Nigam, Priyanka and Yin, Bing},
  journal={arXiv preprint arXiv:2606.12608},
  year={2026}
}