PHYRD - Discover Tool Assisted - Overall WAcc: leaderboard
PHYRD v1:63 Reproduce and43 Discover physics problems; complete problem-set denominators. SAcc averages each problem's correct-subquestion fraction, WAcc requires all subquestions correct. Missing/incomplete conclusions score0. Fields retain their native subsets. GPT-5.6 Sol uses a conversational interface with xhigh effort, distinct from four API models. Source-reported model identities retained without inventing dated endpoints. Discover Tool Assisted. API thinking enabled; Python/SymPy/NumPy/mpmath and public arXiv search/reading,24assistant turns,128toolcalls and2hours/problem. Completion request DeepSeek393216,GLM/Qwen131072,Kimi262144; Qwen thinking262144. Initial109answers plus63continued conversations; all43scored permodel. GPT uses its conversational tool interface. Effort levels are not inferred from enabled thinking. Overall; WAcc. Metric: Accuracy (%).
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Around January 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (xHigh) | 79.07 | #20 (GPT-5.6 Sol) |
| 2 | DeepSeek V4 Pro | 76.74 | #109 |
| 3 | Qwen 3.7 Max | 39.53 | #81 |
| 4 | Kimi K2.6 | 37.21 | #112 |
| 5 | GLM-5.2 | 30.23 | #78 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #6 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=phyrd-discover-tool-assisted-overall-wacc · How It Works · Data refreshed daily, snapshot 2026-10-27.