Reasoning Models and Planning - with Rao Kambhampati (Arizona State)

Rao Kambhampati, professor of computer science at Arizona State University and a former AAAI president, examines what reasoning models do, when they work and when they fail. Drawing on decades of planning research, he defines reasoning and planning, discusses what changed when systems such as o1 and DeepSeek R1 moved verification from inference into post-training, and explains where these models generalise and why benchmarks can mislead. Much of the episode concerns what intermediate chain-of-thought tokens actually do and how outcome-based reinforcement learning affects their meaning, along with world models and agentic safety.