TheoPhysEval - Isolated: leaderboard
TheoPhysEval evaluates autonomous agents on the same 50 research-level theoretical physics problems as its realistic track, across 11 themes. This isolated track permits local numerical or symbolic tools through Harbor but blocks internet access except for the model provider. Scores are the percentage of problems solved in one run, graded against final answers, primarily numerical; no human hints are supplied. Tasks receive two CPUs, 4 GB RAM and eight hours, with 20 hours for the multi-task electronic-structure problems and no token or cost cap. Table 4 of the October 2026 paper reports five agents; model identities retain their harness and reasoning effort. The realistic track is reported separately.
Metric: Problems Solved (%). Source: arxiv.org. Saturation forecast: Around January 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (Max) | 92 | #5 (GPT-6 Astra) |
| 2 | Claude Opus 5.5 (Max) | 92 | #3 (Claude Opus 5.5) |
| 3 | Gemini 3.8 Flash (High) | 62 | #21 (Gemini 3.8 Flash) |
| 4 | GLM-5.3 (Max) | 52 | #40 (GLM-5.3) |
| 5 | DeepSeek V4.1 Flash (Max) | 50 | #36 (DeepSeek V4.1 Flash) |
No result here: #7 Gemini 4 Argon, #8 GPT-6.1 Sol, #9 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=theophyseval-isolated · How It Works · Data refreshed daily, snapshot 2026-10-16.