TAF-MED - Collapse Rate: leaderboard
TAF-MED (arXiv 2608.10258): of the conversations whose first response held the medication-safety boundary (strictly Safe), the share (%) that turn Unsafe at the second or third turn under seven fixed reframes, such as educational, hypothetical, third-person, pharmacy or dose framings. Each model's denominator is its own Safe first turns. Lower is better; over-refusal is not measured. Source: https://arxiv.org/abs/2608.10258
Metric: Collapse rate (%; share of the conversations with a strictly Safe first response that turn Unsafe at the second or third turn; the denominator is each model's own Safe first turns; lower is better). Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Grok 4.3 | 24.4 | #186 |
| 2 | Claude Sonnet 4.6 | 38.8 | #90 |
| 3 | Claude Opus 4.6 | 48.5 | #62 |
| 4 | GPT-5.4 | 67.4 | #78 |
| 5 | GPT-5.4 Mini | 70.9 | #208 |
| 6 | Llama 4 Maverick | 79.5 | #467 |
| 7 | DeepSeek V4 Pro | 81.9 | #100 |
| 8 | Gemini 2.5 Pro | 96.2 | #151 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=taf-med-collapse-rate · How It Works · Data refreshed daily, snapshot 2026-10-19.