Study: Reasoning Model CoT Is Far More Steerable Than Assumed, Can Bypass Monitoring

A new paper by Uriel Dolev et al. finds that reasoning models' chains of thought are far more steerable than current evaluations suggest, and that this steerability could be exploited to bypass monitoring.

2026-11-04 ~ 2026-11-05 · 2 related posts

• New paper: reasoning models' CoT is far more controllable than evals suggest, usable to evade monitors — LChoshen · 2026-11-04
• CoT controllability far higher than evals suggest: self-instructions beat prompts, can evade monitors — kastnerkyle · 2026-11-05