若有疑問，請聯繫公告聯絡人:

計畫助理 林筱萍
國立陽明交通大學 電腦視覺研發中心
TEL：(03)5712121分機52991
FAX：(03)5730785
Add：300093 新竹市大學路1001號 電子資訊大樓608室
-----------------------------------------------------------------------------------

Speaker: Kuang-Huei Lee 李冠輝 (Senior Staff Research Scientist, Google DeepMind)

Time: 2026/05/26 (Tuesday), 13:30-14:30

Location: EC329 (工程三館329室)

 

Title: On Why LLMs are General Problem Solvers and the Path to Physical World

 

Abstract: 

This talk will first discuss how general problem-solving capabilities of Large Language Models (LLMs) emerge from three factors: effective compression of very large data that makes generalization happen; post-training that shapes output computation programs; inference-time computation – what people refer as “thinking” nowadays – that allows solving problems that are more difficult than constant-time complexity. In the second part, we will discuss the path for LLMs to become driving forces of physical AI, solving problems in the real world. In particular, we highlight the critical roles of compositional abstraction and world models.

 

Short Bio:

Kuang-Huei Lee is a Senior Staff Research Scientist at Google DeepMind. His research interests span across generative AI, robotics and reinforcement learning. He notably led RL post-training for the Veo video generation model, co-led AI for chip placement, and co-initiated the RT-1 robotic transformer effort. His research work is widely published in top-tier conferences in machine learning (NeurIPS, ICML, ICLR), robotics (RSS, CoRL, IROS), computer vision (CVPR, ECCV), and NLP (EMNLP). Kuang-Huei obtained his graduate degree from Carnegie Mellon University and his undergraduate degree from National Taiwan University.