SciLaws-Bench: leaderboard
Given real observational data, recover the law behind it — and then do it again in a parallel world where the constants were changed, so a remembered textbook equation is worth nothing. This is the numeric-fit score over all 118 problems; the frontier sits near 50 percent and the gap between single-law and multi-law problems is where it loses most of that.
Metric: S_N Overall (%). Source: yiyihum.github.io. Saturation forecast: Around 2030. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra | 51.33 | #6 |
| 2 | GPT-5.5 | 50.11 | #31 |
| 3 | Claude Opus 5.5 | 48.87 | #3 |
| 4 | Kimi K3 | 47.36 | #30 |
| 5 | Gemini 3.5 Flash | 46.08 | #60 |
| 6 | GLM-5.2 | 45.84 | #73 |
| 7 | Claude Opus 4.8 | 44.94 | #37 |
| 8 | DeepSeek V4 Pro | 43.96 | #102 |
| 9 | Qwen 3.7 Max | 43.8 | #75 |
| 10 | GPT-5 Mini | 40.69 | #186 |
| 11 | DeepSeek V4.1 Flash | 40.59 | #36 |
| 12 | GPT-5.6 Sol | 40.11 | #19 |
| 13 | GPT-5.4 Mini | 37.17 | #212 |
| 14 | GPT-4o Mini | 20.59 | #599 |
No result here: #7 GPT-6.1 Sol, #8 Gemini 4 Argon, #9 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=scilaws-bench · How It Works · Data refreshed daily, snapshot 2026-10-20.