MedPRESS - Conversation Failure 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 conversations with at least one unsafe-agreement answer in the five turns, percent. Lower is better.
Metric: Conversations with any unsafe agreement (%). Source: arxiv.org. Saturation forecast: Not forecast. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 Mini | 50.3 | #228 |
| 2 | Llama 3.3 70B Instruct | 63.5 | #541 |
| 3 | Llama 3.1 70B Instruct | 71.2 | #567 |
| 4 | Llama 3.2 3B Instruct | 72.3 | #1327 |
| 5 | GPT-OSS-120B (Low) | 78.3 | #346 (GPT-OSS-120B) |
| 6 | Llama 3.1 8B Instruct | 86.2 | #1017 |
| 7 | MedGemma-4B-IT | 90.4 | #811 |
| 8 | Qwen 3 14B (Non-reasoning) | 90.5 | #542 (Qwen 3 14B) |
| 9 | GPT-OSS-20B (Low) | 91.6 | #523 (GPT-OSS-20B) |
| 10 | Phi-4 Mini Instruct | 91.7 | #1232 |
| 11 | Qwen 3 8B (Non-reasoning) | 92.9 | #678 (Qwen 3 8B) |
| 12 | Qwen 3 4B (Non-reasoning) | 93.2 | #834 (Qwen 3 4B) |
| 13 | Qwen 3 32B (Non-reasoning) | 94.7 | #449 (Qwen 3 32B) |
| 14 | DeepSeek V4 Flash (Non-reasoning) | 94.7 | #155 (DeepSeek V4 Flash) |
| 15 | Phi-4 | 94.9 | #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-conversation-failure-rate · How It Works · Data refreshed daily, snapshot 2026-11-05.