InfiniteScienceGym: leaderboard
InfiniteScienceGym (2026) procedurally generates scientific data repositories and asks templated questions, answerable or not, whose ground truth comes from the simulator; models browse the files and run Python through tools. Accuracy on 500 templated questions in percent; higher is better.
Metric: Accuracy (%) on 500 templated questions (answerable and unanswerable) sampled from the first 500 simulated repositories, graded against the simulator's ground truth; models read files and run Python through MCP tools (proprietary models natively, open-weight models in a smolagents ReAct loop); higher is better. Source: arxiv.org. Saturation forecast: Around March 2028. 5 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | GPT-5.4 | Medium | 49.2 | #135 |
| 2 | Claude Opus 4.6 | Medium | 43.7 | #113 |
| 3 | GPT-OSS-20B | 29.7 | #1183 | |
| 4 | Qwen 3 4B 2507 Instruct | 26.5 | #2221 | |
| 5 | Gemma 3 27B (IT) | 24.2 | #1324 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=infinitesciencegym · How It Works · Data refreshed daily, snapshot 2026-11-21.