Cave-Bench - Over-Correction Rate: leaderboard
Share of Cave-Bench runs in which the agent, after a false accusation, changed its verified correct work so that replay shows damage. This counts outcomes only, unlike the headline, which also scores confession and evidence use.
Metric: Runs that damaged correct work (%, lower is better). Source: arxiv.org. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 5 | 12.5 | #111 |
| 2 | Qwen 3.8 Max | 17.6 | #55 |
| 3 | Claude Opus 5 | 19.66 | #22 |
| 4 | GLM-5.2 | 21.79 | #138 |
| 5 | Kimi K3 | 23.4 | #54 |
| 6 | Grok 4.5 | 27.3 | #82 |
| 7 | Hy3 | 30.7 | #195 |
| 8 | Qwen 3.7 Plus | 37.74 | #207 |
| 9 | MiniMax-M3 | 41.69 | #269 |
| 10 | GPT-5.6 Sol | 48.2 | #30 |
| 11 | Kimi K2.7 Code | 52.21 | #185 |
| 12 | Qwen 3.5 Flash | 54.34 | #435 |
| 13 | Qwen 3.5 9B | 60 | #803 |
| 14 | MiniMax-M2.7 | 60.06 | #518 |
Interactive version: theaggregate.ai/benchmark?slug=cave-bench-over-correction-rate · How It Works · Data refreshed daily, snapshot 2026-11-13.