MedPRESS - Turn of Flip: 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. Pressure turns withstood before the first unsafe agreement, averaged over conversations: 0 means unsafe at turn 1 and 5 means never. Higher is better.
Metric: Pressure turns withstood before first unsafe agreement (0-5). Source: arxiv.org. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 Mini | 3.37 | #228 |
| 2 | Llama 3.3 70B Instruct | 2.85 | #541 |
| 3 | GPT-OSS-120B (Low) | 2.73 | #346 (GPT-OSS-120B) |
| 4 | Llama 3.1 70B Instruct | 2.64 | #567 |
| 5 | Llama 3.2 3B Instruct | 2.63 | #1327 |
| 6 | Llama 3.1 8B Instruct | 2.6 | #1017 |
| 7 | Phi-4 Mini Instruct | 1.99 | #1232 |
| 8 | MedGemma-4B-IT | 1.98 | #811 |
| 9 | Qwen 3 14B (Non-reasoning) | 1.96 | #542 (Qwen 3 14B) |
| 10 | Qwen 3 8B (Non-reasoning) | 1.87 | #678 (Qwen 3 8B) |
| 11 | GPT-OSS-20B (Low) | 1.74 | #523 (GPT-OSS-20B) |
| 12 | Qwen 3 4B (Non-reasoning) | 1.74 | #834 (Qwen 3 4B) |
| 13 | Phi-4 | 1.57 | #707 |
| 14 | Qwen 3 32B (Non-reasoning) | 1.57 | #449 (Qwen 3 32B) |
| 15 | Qwen 2.5 72B Instruct | 1.55 | #469 |
Interactive version: theaggregate.ai/benchmark?slug=medpress-turn-of-flip · How It Works · Data refreshed daily, snapshot 2026-11-05.