Title: Theory of Mind Reasoning for LLMs: What do we know and how can we make it better?

Speaker: Prof. Hyunwoo Kim @ KAIST

Time : 13:30 - 14:30, Dec 11th, 2025

Location: Hybrid

In-person: 22410

Online: https://hli.skku.edu/InvitedTalk251211 

Language: English speech & English slides

Abstract: 

Reasoning models are excelling in domains like math and coding where obtaining verifiable ground-truth answers is straightforward. However, acquiring such annotated answers for social reasoning is much more challenging. In this talk, I will discuss our recent works examining what these models are missing and how we can improve LLMs by leveraging their existing strengths. Our approach highlights the importance of inferring perception in social reasoning and reveals parallels between social reasoning and Sequential Monte Carlo, a classical machine learning algorithm. We demonstrate that it is possible to generate much more accurate and efficient reasoning trajectories compared to recent reasoning models. Finally, I will share my perspective on the future of theory-of-mind and social reasoning in the rapidly evolving AI research landscape.

Bio:

Hyunwoo Kim is an incoming Assistant Professor at KAIST AI, where he will lead the Computation & Cognition Lab. Currently, he is a postdoctoral scholar working with Yejin Choi at NVIDIA for his gap year. He is interested in building machines capable of reasoning and cognition, by taking an interdisciplinary approach by connecting language models and cognitive science. Hyunwoo has presented his work in top-tier NLP and AI conferences, and he is a recipient of the EMNLP outstanding paper award in 2023. He received his Ph.D. in Computer Science at Seoul National University advised by Gunhee Kim and B.S. in Psychology & Computer Science at Yonsei University in Korea.