ExplorationBench - AlienLogic: leaderboard
AlienLogic, a sandbox of ExplorationBench (Fudan University, Tencent Hunyuan and Tsinghua University, 2026): a first-order natural-deduction system with 24 patched inference rules. The system explores for four rounds by submitting proofs to a checker, then answers 70 held-out tasks, with unprovable theorems credited only when it declines; percent correct of the best of three independent trajectories, each task answered three times, at the highest reasoning setting each API offers, higher is better.
Metric: Held-out accuracy after exploration (%; Best@3: the highest round-4 score among three independent autonomous exploration trajectories of 12 proofs per round; AlienLogic, a first-order natural-deduction system with 24 patched inference rules, 70 tasks checked by a proof checker, designated unprovable ones credited only when the system declines to prove them; 70 held-out tasks each answered three times in a tool-disabled copy of the conversation and averaged; highest reasoning setting each API offers). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Grok 4.6 (xHigh) | 83.8 | #44 (Grok 4.6) |
| 2 | Claude Opus 5 (Max) | 77.6 | #20 (Claude Opus 5) |
| 3 | GPT-5.6 Sol (Max) | 75.2 | #29 (GPT-5.6 Sol) |
| 4 | Hy4 preview (High) | 74.8 | #66 (Hy4 preview) |
| 5 | DeepSeek V4 Pro (Max) | 73.8 | #181 (DeepSeek V4 Pro) |
| 6 | Kimi K3 (Max) | 72.9 | #53 (Kimi K3) |
| 7 | DeepSeek V4.1 Flash (Max) | 72.4 | #58 (DeepSeek V4.1 Flash) |
| 8 | Gemini 3.8 Flash (High) | 58.1 | #36 (Gemini 3.8 Flash) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=explorationbench-alienlogic · How It Works · Data refreshed daily, snapshot 2026-11-12.