Pressure Reveals Character - Corrigibility: Corrigibility: leaderboard
Mean 1-5 judge rubric score over 18 multi-turn scenarios testing accepting valid corrections from users. One behaviour of the Corrigibility category; a subset of the headline's transcripts.
Metric: Mean judge rubric score over 18 scenarios (1 = failure, 5 = exemplary). Source: storage.googleapis.com. 24 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 235B A22B | 4.94 | #623 |
| 2 | GPT-5 Mini | 4.89 | #348 |
| 3 | Gemini 2.5 Flash | 4.83 | #453 |
| 4 | Claude Sonnet 4.5 | 4.83 | #250 |
| 5 | GPT-OSS-120B | 4.78 | #700 |
| 6 | Claude Sonnet 4 | 4.78 | #400 |
| 7 | Mistral Medium 3.1 | 4.72 | #852 |
| 8 | GPT-5.2 | 4.67 | #176 |
| 9 | DeepSeek V3.2 | 4.67 | #406 |
| 10 | Kimi K2 | 4.67 | #439 |
| 11 | Claude Opus 4.5 | 4.67 | #156 |
| 12 | Gemini 3 Flash | 4.61 | #164 |
| 13 | Llama 4 Maverick | 4.56 | #1099 |
| 14 | Qwen 3 14B | 4.56 | #1233 |
| 15 | Llama 3.1 405B Instruct | 4.56 | #1115 |
Interactive version: theaggregate.ai/benchmark?slug=pressure-reveals-character-corrigibility-corrigibility · How It Works · Data refreshed daily, snapshot 2026-11-13.