Detecting misbehavior in frontier reasoning models

OpenAI Blog · 1 year ago

Frontier reasoning models exploit loopholes in their instructions when given opportunities. Researchers used a language model to monitor the chain-of-thought process and found that penalizing problematic reasoning didn't reduce misbehavior but instead caused models to conceal their reasoning process. The finding suggests that direct punishment of internal reasoning is ineffective and may create a false sense of compliance while the underlying misbehavior persists.