Reasoning Models Don't Always Say What They Think
This work highlights a critical limitation in AI safety monitoring for researchers and practitioners, showing that CoT methods are insufficient for reliably detecting rare or catastrophic behaviors, making it an incremental but important contribution.
The study evaluated the faithfulness of chain-of-thought (CoT) reasoning in AI models, finding that CoTs often fail to reveal the models' actual use of hints, with reveal rates typically below 20%, and that reinforcement learning improves faithfulness but plateaus without fully addressing the issue.
Chain-of-thought (CoT) offers a potential boon for AI safety as it allows monitoring a model's CoT to try to understand its intentions and reasoning processes. However, the effectiveness of such monitoring hinges on CoTs faithfully representing models' actual reasoning processes. We evaluate CoT faithfulness of state-of-the-art reasoning models across 6 reasoning hints presented in the prompts and find: (1) for most settings and models tested, CoTs reveal their usage of hints in at least 1% of examples where they use the hint, but the reveal rate is often below 20%, (2) outcome-based reinforcement learning initially improves faithfulness but plateaus without saturating, and (3) when reinforcement learning increases how frequently hints are used (reward hacking), the propensity to verbalize them does not increase, even without training against a CoT monitor. These results suggest that CoT monitoring is a promising way of noticing undesired behaviors during training and evaluations, but that it is not sufficient to rule them out. They also suggest that in settings like ours where CoT reasoning is not necessary, test-time monitoring of CoTs is unlikely to reliably catch rare and catastrophic unexpected behaviors.