Evaluating chain-of-thought monitorability

OpenAI Blog · 7 months ago

OpenAI developed a framework and evaluation suite to assess whether a model's internal reasoning can be effectively monitored during chain-of-thought processes. The framework includes 13 evaluations across 24 different environments to test monitoring capabilities. Monitoring a model's intermediate reasoning steps proved substantially more effective than checking only final outputs, suggesting a potential approach for controlling increasingly capable AI systems.