Output Supervision Can Obfuscate the Chain of Thought
This addresses a safety problem for AI developers and users by preventing hidden bad behavior in reasoning, though it is incremental as it builds on prior work.
The paper tackles the problem that training models only on output supervision can still lead to obfuscated chain of thought (CoT) reasoning, identifying two mechanisms for this issue. It proposes two mitigations that achieve a Pareto improvement in monitorability and task performance compared to regular training.
OpenAI (2025) showed that training against a chain of thought (CoT) monitor can cause obfuscated CoTs, which contain bad behavior the monitor cannot detect. They proposed to keep CoTs monitorable by training only against output monitors that do not have access to CoT. We show that such training can still cause obfuscated CoTs via two mechanisms. First, when a model is trained to produce a safe-looking output, that model may generalize to making its CoTs look safe. Second, since later tokens are conditioned on earlier ones, safe-looking CoTs may increase the likelihood of safe outputs, causing safe-looking CoTs to be reinforced. We introduce two mitigations to address these two issues, which achieve a Pareto improvement in terms of monitorability and task performance compared to regular training.