Special Talk Series — LLMs, Embodied AI, Hybrid Planning, and more (4 lectures)

A four-lecture special talk series by Prof. Scott Sanner (University of Toronto), held 5/26–5/29, 2025.

Lecture 1 — Verifiable, debuggable, and repairable formal reasoning with LLMs. LLMs are susceptible to reasoning errors and hallucinations. We present LLM-TRes, a logical reasoning framework based on “theory resolution” that integrates commonsense knowledge from LLMs with a verifiable logical reasoning framework that mitigates hallucinations and facilitates debugging and repair.

Lecture 2 — Embodied AI Planning via regression in partially observed environments. We introduce lifted regression for planning and extend it to LLM-Regress, an open-world planning approach that integrates lifted regression with LLM-generated affordances to produce sound and complete plans in compact lifted form.

Lecture 3 — Hybrid Traditional+Modern Planning and RL methods for MDPs. Covers the Relational Dynamic Influence Diagram Language (RDDL), a range of planning methodologies (MCTS, mathematical programming, gradient-based optimization, symbolic methods), and a hands-on session using the JaxPlan gradient-based planner.

Lecture 4 — Opportunities and advances in conversational recommendation with generative AI. The need for personalized recommendation in conversational AI assistants, recent developments up to modern LLM-enhanced (agentic) conversational recommendation, and the technical and societal challenges of this shift.

Speaker

Scott Sanner is a Professor in Industrial Engineering and cross-appointed in Computer Science at the University of Toronto. His research spans sequential decision-making, (conversational) recommender systems, and machine/deep learning.