CoT-Control - Information Suppression: leaderboard
A reasoning model gets a question with an instruction, appended to the user message, to keep certain words, or the whole question, out of its chain of thought; the number is how often its reasoning obeys. Lower means the reasoning is harder to bend, which removes one way a model could defeat a monitor that reads it; the opposite reading of an instruction-following score such as IFEval.
Metric: CoT controllability (% of compliant reasoning traces, lower is safer). Source: arxiv.org. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek R1 | 0.2 | #473 |
| 2 | GPT-5 | 0.3 | #167 |
| 3 | O4 Mini | 1.1 | #353 |
| 4 | Qwen 3 14B (Reasoning) | 1.8 | #1233 (Qwen 3 14B) |
| 5 | O3 | 2.1 | #228 |
| 6 | GPT-OSS-20B | 2.5 | #1147 |
| 7 | Qwen 3 32B (Thinking) | 5.3 | #960 (Qwen 3 32B) |
| 8 | GPT-5.1 | 5.6 | #229 |
| 9 | GPT-OSS-120B | 14 | #700 |
| 10 | Claude 3.7 Sonnet (Thinking) | 15 | #458 (Claude 3.7 Sonnet) |
| 11 | Claude Sonnet 4.5 (Thinking) | 15.6 | #250 (Claude Sonnet 4.5) |
| 12 | Claude Sonnet 4 (Thinking) | 18.6 | #400 (Claude Sonnet 4) |
Interactive version: theaggregate.ai/benchmark?slug=cot-control-information-suppression · How It Works · Data refreshed daily, snapshot 2026-11-13.