Study Finds Chain-of-Thought Controllability Can Undermine Monitorability
A new arXiv paper argues that chain-of-thought monitoring, used to detect hidden goals and off-task behavior in AI agents, carries a risk: making CoT more controllable can reduce how well it can be monitored.
Read the full story at arXiv cs.AI ↗
Timeline · 1 report
- 2026-11-18 05:00 · arXiv cs.AI
Chain-of-Thought Controllability Can Reduce Chain-of-Thought Monitorability