HyperLogic: A Hard, Forward-Authored Chinese Logical Reasoning Benchmark with Execution-Derived Answers

📄 Paper  |   💻 GitHub  |   📮 Evaluate on the Full Benchmark

HyperLogic is a Chinese logical-reasoning benchmark built forward: problems are written first, and their answers are derived afterwards by executing formal models of them. A multi-agent workflow hardens seeds written by logic-trained undergraduates without solving them; two agents from different model families independently translate each finished item into executable finite-domain models, and logic solvers derive the answers. An item is released only when its two independently corrected encodings agree, after layered, agent-assisted adjudication under human-expert oversight.

This dataset is the public 30% of HyperLogic: a fixed, uniformly sampled 30% of the items of every tier, with gold answers. The remaining 70% and the formal translations of all items are held out so that results on them stay free of contamination. The prompts, scorer, and Hard agent toolkit are in the GitHub repository.

Configurations

Config	Tier	Items	Sub-questions	Full tier
seed	HyperLogic-Seed	69	69	228 items
base	HyperLogic-Base	59	274	195 items · 922 sub-questions
hard	HyperLogic-Hard	30	30	100 items

• Seed: single-question seeds written by undergraduates with coursework in logic.
• Base: a shared background (about 3,300 Chinese characters) with 4.7 sub-questions on average. An item counts as correct only if every sub-question is.
• Hard: single questions over larger, coupled search spaces (backgrounds of about 8,700 characters), intended for direct answering and for agents that write code in a sandbox.

Samples are item-level: all sub-questions of a selected Base item are kept together.

Usage

from datasets import load_dataset

seed = load_dataset("tencent/HyperLogic", "seed", split="test")
base = load_dataset("tencent/HyperLogic", "base", split="test")
hard = load_dataset("tencent/HyperLogic", "hard", split="test")

Answers are free-form Chinese text and are scored with a semantic-equivalence LLM judge, not by string matching. To reproduce the paper's protocol, use evaluate.py from the GitHub repository, which writes the direct-answer prompts and scores predictions; the Hard agent conditions (sandbox only, and sandbox with the logic modeling library) are in its hard_agent/ folder.

Data Fields

Field	Type	Description
id	string (Seed) / int (Base, Hard)	Item ID. Seed IDs are content-derived seed-... strings.
title	string	Short Chinese title (Seed and Base only).
background	string	The scenario, in Chinese.
question	list of strings	The question(s); one for Seed and Hard.
answer	list of strings	Gold answers, aligned with question.

Leaderboard

Accuracy (%) on the full benchmark, mean ± standard deviation over three rollouts, every model at its highest reasoning setting. Scores on this public subset are not comparable to these numbers.

Model	Seed	Base Item	Base Sub-Q	Hard Direct	Hard Sandbox	Hard + Library
GPT-5.6 Sol	89.8 ± 3.7	77.6 ± 0.6	92.6 ± 0.2	13.7 ± 5.9	40.3 ± 4.2	45.7 ± 2.1
Claude Opus 5	86.5 ± 0.7	74.2† ± 2.7	88.9† ± 0.7	15.6† ± 1.8	55.7 ± 3.2	57.0 ± 2.6
Kimi K3	88.6 ± 0.0	57.9 ± 1.8	81.6 ± 1.8	4.0 ± 0.0	41.3 ± 1.5	43.0 ± 8.9
Grok 4.6	87.0 ± 1.8	52.5 ± 4.1	82.9 ± 0.7	12.0 ± 1.0	39.3 ± 1.5	38.3 ± 2.1
DeepSeek V4 Pro	83.9 ± 1.7	51.8 ± 1.5	81.2 ± 1.2	9.3 ± 1.2	26.0 ± 5.6	20.0 ± 1.7
GLM-5.3	87.3 ± 0.9	45.6 ± 3.2	67.2 ± 4.9	3.0 ± 1.0	20.7 ± 0.6	31.3 ± 5.1
Doubao Seed 2.1 Pro	85.8 ± 0.7	44.6 ± 2.2	73.2 ± 1.0	5.0 ± 1.0	26.0 ± 1.0	33.7 ± 0.6

^†Excludes items refused by the model's API service. Grok 4.6 runs at xhigh, every other model at max.

Selection

Each tier is sampled independently and uniformly without replacement with random.Random("2027:<tier>"), releasing ceil(0.30 × full item count) items from the full corpus sorted by item ID. The selection script and SHA-256 commitments to every full tier are in the GitHub repository (build_public_subset.py, selection_manifest.json).

Evaluate on the Full Benchmark

The held-out 70% is not public. To have a model evaluated on the full benchmark, email Ming Zhang (mingzhang23@m.fudan.edu.cn) or Maxm Pan (maxmpan@tencent.com), and we will run it under the same protocol as the leaderboard.

License

Apache-2.0.

Citation

@article{zhang2026hyperlogic,
  title   = {HyperLogic: A Hard, Forward-Authored Chinese Logical Reasoning
             Benchmark with Execution-Derived Answers},
  author  = {Zhang, Ming and Peng, Qiyuan and Wei, Yinxi and Shen, Yujiong and
             Tan, Kexin and Wang, Yuhui and Xiang, Zhenghao and Ye, Junjie and
             Yin, Zhangyue and Xi, Zhiheng and Dou, Shihan and Wang, Weikang and
             Zhang, Yuhao and Gui, Tao and Yang, Ruizhi and Zhang, Qi and
             Huang, Xuanjing and Chen, Alex and Pan, Maxm},
  journal = {arXiv preprint arXiv:2605.19597},
  year    = {2026}
}