[107th TrustML Young Scientist Seminar] Talk by Prof. Mengyue Yang (University of Bristol) "Grounded LLM Reasoning: From Causal Reasoning to World Models"

Date and Time: Nov 13, 2026, 13:30 -- 14:30 (JST)
Venue: Hybrid (Online + Meeting Room C, RIKEN Nihonbashi Office)
*In-person attendance is limited to AIP researchers.

Title: Grounded LLM Reasoning: From Causal Reasoning to World Models

Speaker:
Prof. Mengyue Yang (University of Bristol)

Abstract:
Large language models can produce convincing reasoning traces, but when does this reasoning reflect the real understanding? This talk explores how can causality help ground LLM reasoning and the reasoning improved itself. I will first introduce causal sufficiency and necessity as criteria for evaluating chain-of-thought reasoning, showing how they can identify critical reasoning steps and improve both accuracy and efficiency. I will then discuss the limitations of outcome-based reinforcement learning and how natural-language feedback can support learning from experience. Finally, I will turn to world models, which connect learned representations with causal structure to support intervention, counterfactual reasoning, and planning. Through recent work on causality-seeking agents, I will discuss how agents can actively explore their environments and continually revise their causal understanding. Together, these directions point towards agents whose reasoning is grounded in the mechanisms and consequences of their interactions with the world.

Short Bio:
Dr. Mengyue Yang is a Lecturer (Assistant Professor) at the University of Bristol. She received her PhD from the UCL Centre for Artificial Intelligence. Her research focuses on world models, causality, and reinforcement learning. She was named a KAUST Rising Star in AI in 2024 and received the AAAI 2026 New Faculty Highlights Award. She serves as an Area Chair for NeurIPS and a Guest Editor for a special issue of the Machine Learning Journal. She has also served as an organizer and PC for workshops at ICLR 2024–2025 and the NeurIPS 2025 Workshop on Embodied World Models.