MedDialBench - Occasional Contradiction: leaderboard

MedDialBench (2026) has doctor-agent LLMs interview a simulated patient (Claude Opus 4.5 following case-specific behaviour scripts) on 85 OSCE cases and name a diagnosis, which Qwen3-Max judges correct or not. This board uses a patient who occasionally contradicts themselves (logic, moderate). Diagnostic accuracy in percent; higher is better.

Metric: Diagnostic accuracy (%) over 85 OSCE cases (each cell counts whole cases out of 85) after a free-form consultation with a patient who occasionally contradicts themselves (logic, moderate), the final diagnosis judged correct or not by Qwen3-Max; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 3.1 Pro (Preview)85.9#99
2GPT-5.4Non-reasoning78.8#135
3Claude Opus 4.677.6#113
4Qwen 3.5 PlusNon-reasoning70.6#242
5DeepSeek V3.2Non-reasoning65.9#419

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

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