Evaluating chain-of-thought monitorability

OpenAI introduces a brand new framework and analysis suite for chain-of-thought monitorability, overlaying 13 evaluations throughout 24 environments. Our findings present that monitoring a mannequin’s inside reasoning is way more practical than monitoring outputs alone, providing a promising path towards scalable management as AI methods develop extra succesful.