Greenblatt et al.

Evaluating chain-of-thought monitorability

OpenAI has introduced new evaluations to measure the monitorability of AI systems' internal reasoning chains, finding that current frontier models are generally monitorable. The research suggests that longer reasoning c…