MedPRESS - Safe Stance Adherence 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 keep the safe medical stance, percent. Higher is better.
Metric: Safe-stance answers (%; turns 1-5, LLM judge). Source: arxiv.org. Saturation forecast: Around March 2027. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 Mini | 74.7 | #228 |
| 2 | GPT-OSS-120B (Low) | 47.7 | #346 (GPT-OSS-120B) |
| 3 | GPT-OSS-20B (Low) | 32.4 | #523 (GPT-OSS-20B) |
| 4 | Llama 3.1 70B Instruct | 31.1 | #567 |
| 5 | Qwen 3 8B (Non-reasoning) | 30.6 | #678 (Qwen 3 8B) |
| 6 | Llama 3.1 8B Instruct | 30.2 | #1017 |
| 7 | DeepSeek V4 Flash (Non-reasoning) | 30 | #155 (DeepSeek V4 Flash) |
| 8 | Qwen 3 14B (Non-reasoning) | 29.5 | #542 (Qwen 3 14B) |
| 9 | Llama 3.3 70B Instruct | 29 | #541 |
| 10 | Llama 3.2 3B Instruct | 25 | #1327 |
| 11 | Gemma 3 27B (IT) | 24.4 | #528 |
| 12 | Qwen 2.5 72B Instruct | 24.4 | #469 |
| 13 | Qwen 3 4B (Non-reasoning) | 24 | #834 (Qwen 3 4B) |
| 14 | Qwen 3 32B (Non-reasoning) | 23.3 | #449 (Qwen 3 32B) |
| 15 | Phi-4 | 20.6 | #707 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=medpress-safe-stance-adherence-rate · How It Works · Data refreshed daily, snapshot 2026-11-05.