Host: 周源

Abstract:

Planning is a crucial element of both human intelligence and contemporary large language models (LLMs). In this talk, I introduce the project ALPINE, which initiates a theoretical investigation into the emergence of planning capabilities in Transformer-based autoregressive LLMs. We model planning as a network path-finding task, where the objective is to generate a valid path from a specified source node to a designated target node. Our gradient-based theoretical analysis of learning dynamics reveals that LLMs can learn both the adjacency and a limited form of the reachability matrices, which are validated by experiments. Our analysis uncovers a limitation of the autoregressive Transformer architecture --- it cannot identify transitive reachability relationships. We further investigate the impact of reinforcement learning with verifiable reward (RLVR) and invalid transition detection together with retry harness on language model planning. Our findings provide new insights into how the internal mechanisms of autoregressive learning facilitate intelligent planning and deepen our understanding of how future LLMs might achieve more advanced and general planning-and-reasoning capabilities across diverse applications.

Bio:

Wei Chen is a Principal Researcher at Microsoft Research Asia, and the Chair of MSRA Theory Center. He has served as a guest professor at several universities including Tsinghua University, Shanghai Jiao Tong University, Hong Kong University of Science and Technology – Guangzhou, and Shenzhen University. He is a Fellow of the Institute of Electrical and Electronic Engineers (IEEE) and the Association of Computing Machinery (ACM). His main research interests include foundations of language models, online learning and optimization, social and information networks, network game theory and economics, distributed computing, and fault tolerance.