Refusal Self-Prediction - Harmful Requests: leaderboard
Refusal self-prediction accuracy on requests the model nearly always refuses (empirical refusal rate 80-100%). Higher is better. Source: arxiv.org/abs/2604.00228.
Metric: Prediction Accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4 | 100 | #208 |
| 2 | Llama 3.1 405B | 98.9 | #476 |
| 3 | Claude Sonnet 4.5 | 96.5 | #148 |
| 4 | GPT-5.2 | 95.8 | #114 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #6 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=refusal-self-prediction-harmful-requests · How It Works · Data refreshed daily, snapshot 2026-11-06.