File size: 2,955 Bytes
34d5cb1
 
39af568
 
 
 
 
 
 
 
 
 
 
 
34d5cb1
39af568
 
 
 
 
 
 
 
 
 
 
 
 
 
 
110c756
 
 
39af568
 
 
110c756
39af568
 
 
110c756
39af568
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
---
license: apache-2.0
language:
- en
pretty_name: ComBench
task_categories:
- text-generation
tags:
- mathematics
- combinatorics
- olympiad
- reasoning
- benchmark
paperswithcode_id:
---

# ComBench

ComBench is an Olympiad-level combinatorics benchmark for evaluating rigorous
proof reasoning and constructive realization in large language models. It
contains 100 human-annotated competition-level problems:

- 50 analysis-centric records in `data/analysis.jsonl`
- 50 construction-centric records in `data/construction.jsonl`

Construction-centric records additionally include explicit construction
instructions, reference constructions, and deterministic Python verifiers.

## Usage

Install the Hugging Face `datasets` package and load ComBench directly from the
Hub:

```python
from datasets import load_dataset

dataset = load_dataset("Simplified-Reasoning/ComBench")
print(dataset)
```

The dataset can also be loaded from the raw JSONL files in this repository:

```python
from datasets import load_dataset

dataset = load_dataset(
    "json",
    data_files={
        "analysis": "data/analysis.jsonl",
        "construction": "data/construction.jsonl",
    },
)
```

## Fields

Common fields:

- `id`: problem identifier used by ComBench.
- `split`: `analysis` or `construction`.
- `source`: normalized contest source identifier.
- `category`: primary combinatorics category.
- `query`: problem statement.
- `ref_answer`: reference final answer when a concise final answer is applicable.
- `ref_solution`: reference solution.
- `grading_guidelines`: rubric used for proof grading.

Construction-centric fields:

- `instruction`: explicit construction task.
- `ref_construction`: reference witness.
- `verify_code`: deterministic Python verifier for checking a construction.

Per-problem category metadata is also provided in `metadata/categories.csv`.

## Safety Note

The `verify_code` field is executable Python code. It is released as benchmark
metadata for deterministic construction verification. Execute it only in a
sandboxed environment.

## Links

- Paper: https://arxiv.org/abs/2606.10479
- Project page: https://simplified-reasoning.github.io/ComBench/docs/
- GitHub: https://github.com/Simplified-Reasoning/ComBench
- Hugging Face paper page: https://huggingface.co/papers/2606.10479

## Citation

```bibtex
@misc{combench2026,
  title        = {ComBench: A Benchmark for Rigorous Proof Reasoning and Constructive Realization in Olympiad-Level Combinatorics},
  author       = {Zhang, Shunkai and Zhang, Haoran and Luo, Yun and Cheng, Qianjia and Lei, Haodi and Li, Yizhuo and Zhan, Runzhe and Wang, Zhilin and Xu, Bangjie and Su, Yucheng and Han, Xinmiao and Qu, Xiaoye and Liu, Dongrui and Lin, Zhouchen and Qiao, Yu and Ding, Ning and Li, Yafu and Cheng, Yu},
  year         = {2026},
  eprint       = {2606.10479},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  url          = {https://arxiv.org/abs/2606.10479}
}
```