New paper: standard CoT monitorability metric is only an upper bound, improvable via self-distillation and RL

armancohan · x · 2026-11-03

A new paper from armancohan's team examines whether LLMs can be trained to make their chains-of-thought easier to monitor:

Shows the standard monitorability metric (TPRmc) is only an upper bound and can overestimate causal detection of problematic behavior
Derives a corresponding lower bound, clarifying what monitorability metrics actually tell us
Proposes three post-training methods; experiments show monitorability can be improved via self-distillation and RL training

Bottom line: yes, CoTs can be trained to be more monitorable — but the choice of evaluation metric can change the conclusion.

More from Safety