CoT-Control - Information Addition: leaderboard
Same questions with an appended instruction to insert text into the chain of thought (a marker at start and end, "safe" after each sentence, "meow" between words); the number is how often the reasoning complies. Lower is safer: it measures whether a model could plant misleading content for a monitor, not whether it answers correctly.
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 | #473 |
| 2 | Claude Sonnet 4.5 (Thinking) | 0 | #250 (Claude Sonnet 4.5) |
| 3 | Qwen 3 14B (Reasoning) | 0 | #1233 (Qwen 3 14B) |
| 4 | O4 Mini | 0.8 | #353 |
| 5 | GPT-OSS-20B | 0.9 | #1147 |
| 6 | Qwen 3 32B (Thinking) | 2.1 | #960 (Qwen 3 32B) |
| 7 | GPT-OSS-120B | 4.8 | #700 |
| 8 | O3 | 5.5 | #228 |
| 9 | GPT-5.1 | 6.2 | #229 |
| 10 | GPT-5 | 7.8 | #167 |
| 11 | Claude Sonnet 4 (Thinking) | 9.6 | #400 (Claude Sonnet 4) |
| 12 | Claude 3.7 Sonnet (Thinking) | 49.9 | #458 (Claude 3.7 Sonnet) |
Interactive version: theaggregate.ai/benchmark?slug=cot-control-information-addition · How It Works · Data refreshed daily, snapshot 2026-11-13.