Tencent Hunyuan releases HyperLogic, a hard Chinese logical reasoning benchmark with execution-derived answers
Tencent Hunyuan open-sourced HyperLogic, a Chinese logical reasoning benchmark with 195 Base and 100 Hard items whose answers are derived by executing formal models. The repository and Hugging Face dataset went public with about 30% of items per tier, gold answers, prompts, scorer, and Hard agent evaluation.
China context
- Original name
- 腾讯混元
- Outside China
- Open weights · github.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can use the released benchmark data and evaluation code to test their models' Chinese logical reasoning, but only about 30% of items are currently public.
- For investors
- The benchmark's difficulty and discriminative power may influence which Chinese model providers are perceived as leaders in logical reasoning, potentially affecting enterprise adoption and investment decisions.
Translated from Chinese. Quotes and facts link to the original sources.
HyperLogic is a forward-authored Chinese logical reasoning benchmark where problems are written first and answers are derived afterwards by executing formal models. It includes HyperLogic-Seed (228 items), HyperLogic-Base (195 items with 922 sub-questions), and HyperLogic-Hard (100 items). The benchmark evaluates frontier models via direct answering and, for Hard, sandbox agents with or without a logic modeling library. On Base, model accuracy ranges from 44.6% to 77.6%, with GPT-5.6 Sol leading at 77.6%. The repository and dataset were released with partial data and full Hard agent evaluation.
The benchmark's construction pipeline separates problem authoring from answer generation: a multi-agent workflow hardens Chinese seeds written by logic-trained undergraduates, then two agents from different model families independently translate each item into executable finite-domain models, and logic solvers derive answers. Items are released only when two corrected encodings agree. This allows items to be harder than the models that built them while keeping answers executable and auditable. The Hard tier includes sandbox evaluation with and without a logic modeling library, and the repository provides the direct-answer prompt, scorer, and complete Hard agent evaluation payloads.
Frontier model developers now have a benchmark that separates models on Chinese logical reasoning where existing benchmarks fail to discriminate; on HyperLogic-Base, the seven evaluated models spread over 33.0 points (44.6–77.6%), whereas on Seed they are within 5.9 points. This gives labs a harder evaluation target for Chinese reasoning and may shift training focus toward longer, multi-step logical problems.
For model providers targeting Chinese-language applications, HyperLogic offers a more discriminative evaluation of logical reasoning, potentially influencing model selection and development priorities. The open-source release allows developers to benchmark their own models against frontier systems.
The next verifiable signal is whether the full HyperLogic dataset (all items and answers) is released, and whether additional frontier models are evaluated and added to the leaderboard. Another signal is whether the benchmark is adopted in model development or cited in subsequent model releases.