New work on CoT monitorability!
We show the standard metric (TPR_mc) is only an upper bound and can overestimate causal detection.
We then derive a lower bound, and show monitorability can be improved via self-distillation and RL training.👇
Can we train LLMs to make their CoTs easier to monitor? 🔭Our experiments say yes. But the evaluation metric can change the conclusion.
(1/6) In our new paper, we develop three post-training methods and derive causal bounds to clarify what monitorability metrics actually tell us
