ExplorationBench - AlienCode: leaderboard
AlienCode, a sandbox of ExplorationBench (Fudan University, Tencent Hunyuan and Tsinghua University, 2026): a small calculation language whose familiar operators follow 31 hidden altered semantics that contradict its manual. The system explores for four rounds by running programs, then writes programs for 70 held-out tasks graded on private inputs; percent correct of the best of three independent exploration 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 tool calls per round; AlienCode, a calculation language with 31 hidden altered semantics, 70 program-synthesis and prediction tasks graded by an interpreter on private inputs; 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 | Claude Opus 5 (Max) | 87.6 | #20 (Claude Opus 5) |
| 2 | GPT-5.6 Sol (Max) | 87.1 | #29 (GPT-5.6 Sol) |
| 3 | Kimi K3 (Max) | 77.6 | #53 (Kimi K3) |
| 4 | Gemini 3.8 Flash (High) | 77.6 | #36 (Gemini 3.8 Flash) |
| 5 | Grok 4.6 (xHigh) | 67.6 | #44 (Grok 4.6) |
| 6 | Hy4 preview (High) | 62.4 | #66 (Hy4 preview) |
| 7 | DeepSeek V4.1 Flash (Max) | 58.1 | #58 (DeepSeek V4.1 Flash) |
| 8 | DeepSeek V4 Pro (Max) | 12.9 | #181 (DeepSeek V4 Pro) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=explorationbench-aliencode · How It Works · Data refreshed daily, snapshot 2026-11-12.