Evaluating chain-of-thought monitorability

OpenAI introduces a new framework and evaluation suite for chain-of-thought monitorability, covering 13 evaluations across 24 environments. Our findings show that monitoring a model’s internal reasoni

• •OpenAI introduces a new framework and evaluation suite for chain-of-thought monitorability, covering 13 evaluations acro
• •Our findings show that monitoring a model’s internal reasoning is far more effective than monitoring outputs alone, offe