HyperLogic - Hard (Sandbox Agent): leaderboard
HyperLogic (Tencent Hunyuan and Fudan, arXiv 2605.19597v2; first released as LLMEval-Logic) is a Chinese logical-reasoning set built by forward construction: problems are written in natural language first and formalized afterwards, with each answer derived by two independent formalizations and logic solvers. Unlike reverse-constructed sets such as ZebraLogic, which render a predefined formal structure into text, no formalization is fixed in advance, so "choosing a faithful model is itself part of the reasoning task". Every model runs at its highest reasoning setting (Grok 4.6 xhigh, the others max); free-form Chinese answers are graded by a semantic-equivalence LLM judge, and the score is accuracy (%) on the full set, the mean of three rollouts. This board: the same 100 Hard questions solved by the model as an agent with a code sandbox. Higher is better. Source: https://arxiv.org/abs/2605.19597
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Around December 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 (Max) | 55.7 | #16 (Claude Opus 5) |
| 2 | Kimi K3 (Max) | 41.3 | #35 (Kimi K3) |
| 3 | GPT-5.6 Sol (Max) | 40.3 | #22 (GPT-5.6 Sol) |
| 4 | Grok 4.6 (xHigh) | 39.3 | #29 (Grok 4.6) |
| 5 | DeepSeek V4 Pro (Max) | 26 | #112 (DeepSeek V4 Pro) |
| 6 | GLM-5.3 (Max) | 20.7 | #44 (GLM-5.3) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=hyperlogic-hard-sandbox-agent · How It Works · Data refreshed daily, snapshot 2026-11-05.