Cave-Bench - False Confession Rate: leaderboard
Share of Cave-Bench runs in which the agent accepts responsibility for a failure it did not cause. This scores what the agent says, whereas the over-correction board scores what it does to the work.
Metric: Runs accepting the false blame (%, lower is better). Source: arxiv.org. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 5 | 3.53 | #111 |
| 2 | GLM-5.2 | 7.86 | #138 |
| 3 | Qwen 3.8 Max | 8.66 | #55 |
| 4 | Claude Opus 5 | 8.71 | #22 |
| 5 | Hy3 | 16.96 | #195 |
| 6 | Kimi K3 | 16.99 | #54 |
| 7 | Grok 4.5 | 28.97 | #82 |
| 8 | Qwen 3.7 Plus | 29.2 | #207 |
| 9 | MiniMax-M3 | 33.52 | #269 |
| 10 | GPT-5.6 Sol | 41 | #30 |
| 11 | Kimi K2.7 Code | 44.25 | #185 |
| 12 | MiniMax-M2.7 | 51.84 | #518 |
| 13 | Qwen 3.5 Flash | 58.26 | #435 |
| 14 | Qwen 3.5 9B | 68.73 | #803 |
Interactive version: theaggregate.ai/benchmark?slug=cave-bench-false-confession-rate · How It Works · Data refreshed daily, snapshot 2026-11-13.