DiscoverPhysics - Pass@1: leaderboard

DiscoverPhysics (2026) asks an LLM agent to discover the laws of motion of 22 simulated worlds whose physics deviates from our own by designing experiments, then to submit an explanation and a Python implementation of the law. A world passes when the law predicts held-out trajectories within a normalized MSE of 0.1 and the explanation scores at least 0.9 against an expert rubric; this board is the expected percent of worlds passed in one attempt, higher is better.

Metric: Expected share of the 22 worlds (%) passed in one attempt, estimated from five seeds per world: a world passes when the submitted law predicts held-out test particles with a variance-normalized trajectory MSE below 0.1 and the explanation scores at least 0.9 against the expert rubric (claude-opus-4-6 judge); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 11 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Claude Opus 4.7High26.4#85
2GPT-5.5Medium21.7#54
3Qwen 3.5 397B A17B4.5#264
4Kimi K2.52.6#266
5Claude Sonnet 4.61.8#173
6Qwen 3 235B A22B1#628
7GPT-5.40.9#135
8Llama 3.3 70B Instruct0#1385
9GPT-OSS-120B0#704
10Claude Haiku 4.50#583
11GPT-OSS-20B0#1183

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=discoverphysics-pass-1 · How It Works · Data refreshed daily, snapshot 2026-11-21.