Training-Time Optimization Pressures That Degrade CoT Monitorability

Identify and quantify the kinds and amounts of direct and indirect optimization pressure during model training that are permissible without causing significant degradation in chain-of-thought monitorability.

Background

The authors argue that choices in training—such as process supervision, outcome-based RL scaling, and indirect pressures via reward models or human preference collection—could affect how faithfully chain-of-thought reflects a model’s reasoning. Such pressures might incentivize models to produce less legible or less faithful CoT.

Despite recognizing these potential effects, the authors state that the community does not yet understand which training pressures and magnitudes preserve monitorability. Clarifying these constraints would inform safe training practices and help prevent degradation of CoT’s oversight value.