MedPRESS - Unsafe Agreement Rate: leaderboard
MedPRESS (arXiv 2608.02520v1, Table 3): 600 everyday medical cases run as five-turn conversations in which a simulated patient escalates pressure to accept an unsafe belief, under four prompt conditions (2,400 conversations per model), judged by Qwen3-32B. Share of answers over turns 1-5 that agree with the unsafe belief, percent; ambiguous non-rejections count as neither agreement nor adherence. Lower is better.
Metric: Unsafe-agreement answers (%; turns 1-5, LLM judge). Source: arxiv.org. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 Mini | 21.9 | #228 |
| 2 | Llama 3.3 70B Instruct | 34.7 | #541 |
| 3 | GPT-OSS-120B (Low) | 34.9 | #346 (GPT-OSS-120B) |
| 4 | Llama 3.1 70B Instruct | 41.9 | #567 |
| 5 | Llama 3.2 3B Instruct | 42.8 | #1327 |
| 6 | Llama 3.1 8B Instruct | 43 | #1017 |
| 7 | MedGemma-4B-IT | 46.2 | #811 |
| 8 | Phi-4 Mini Instruct | 46.8 | #1232 |
| 9 | Qwen 3 14B (Non-reasoning) | 50.4 | #542 (Qwen 3 14B) |
| 10 | Qwen 3 8B (Non-reasoning) | 50.8 | #678 (Qwen 3 8B) |
| 11 | GPT-OSS-20B (Low) | 52.4 | #523 (GPT-OSS-20B) |
| 12 | Qwen 3 4B (Non-reasoning) | 54.9 | #834 (Qwen 3 4B) |
| 13 | Phi-4 | 56.1 | #707 |
| 14 | Qwen 3 32B (Non-reasoning) | 57.8 | #449 (Qwen 3 32B) |
| 15 | Gemma 3 27B (IT) | 61.4 | #528 |
Interactive version: theaggregate.ai/benchmark?slug=medpress-unsafe-agreement-rate · How It Works · Data refreshed daily, snapshot 2026-11-05.