MedPRMBench (Generative): leaderboard

Step-level error detection in medical reasoning chains on MedPRMBench (2026), process reward benchmark chains with 14 types of injected and expert-revised reasoning errors, the model writing one correct or erroneous verdict per step. PRMScore, the mean of the F1 for erroneous steps and the F1 for correct steps, in percent; higher is better.

Metric: PRMScore (%): mean of the F1 for erroneous steps and the F1 for correct steps (weights 0.5 and 0.5) in step-level error detection over the erroneous reasoning chains of the 6,500 MedPRMBench test instances (14 injected medical reasoning error types), generative protocol: the API model receives the medical question with numbered reasoning steps and outputs one plus or minus symbol per step; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-5.475.4#135
2GPT-5.274.6#185
3Claude Opus 4.573.7#160
4GLM-4.772.4#380
5DeepSeek V3.170.7#511
6Qwen 3 Max70.4#415
7DeepSeek R167.8#482
8DeepSeek V3.265.4#419
9Gemini 3.1 Pro (Preview)63.6#99

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

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