IsoSci: leaderboard
IsoSci (2026) pairs 288 short procedural science problems across physics, chemistry, biology and earth science so that both members of a pair share one solution structure but need different domain knowledge. Zero-shot chain-of-thought accuracy in percent, higher is better; truncated or empty answers count as wrong.
Metric: Accuracy (%; 288 problems forming 144 isomorphic cross-domain pairs across physics, chemistry, biology and earth science, each pair sharing one solution structure but needing disjoint domain knowledge; zero-shot chain-of-thought, temperature 0, 8,192-token limit, one sample; letter match for four-way multiple choice, 2% relative tolerance for numeric answers; truncated, empty or failed responses count as wrong). Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.0 Flash (001) | 62.2 | #727 |
| 2 | GPT-4o Mini (2024-07-18) | 61.5 | #1510 |
| 3 | Qwen 3 32B (Thinking) | 37.2 | #960 (Qwen 3 32B) |
| 4 | Qwen 3 32B (Non-reasoning) | 36.8 | #960 (Qwen 3 32B) |
| 5 | O3 Mini | 36.8 | #533 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=isosci · How It Works · Data refreshed daily, snapshot 2026-11-13.