PhysDox - End-to-End Diagnosis: leaderboard
PhysDox (2026) tests whether LLMs can audit biomedical physiological-sensing protocols for physical feasibility on a 683-protocol expert-curated Gold set spanning six sensing domains. This board is the end-to-end accuracy (%) on fatal protocols: a protocol counts only if it is first flagged as fatal and then its violated fine-grained constraint is named, under zero-shot prompting; higher is better.
Metric: End-to-end Fine-12 accuracy (%) on the fatal-violation protocols of the 683-protocol expert-curated Gold set of biomedical physiological-sensing protocols (six domains): a protocol counts only if Stage 1 flags it as fatal and Stage 2 names the violated fine-grained constraint; zero-shot prompting, temperature 0.1, JSON output with retries for malformed responses; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 52.5 | #49 |
| 2 | DeepSeek V4 Flash | 49.3 | #258 |
| 3 | Claude Opus 4.7 | 46.7 | #79 |
| 4 | DeepSeek V4 Pro | 39.3 | #188 |
| 5 | MiMo-V2.5-Pro | 27.6 | #279 |
| 6 | MiMo-V2.5 | 26 | #354 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=physdox-end-to-end-diagnosis · How It Works · Data refreshed daily, snapshot 2026-11-13.