Reinforcement Learning for Language Models That Think and Act

PhD Qualifying Examination


Title: "Reinforcement Learning for Language Models That Think and Act"

by

Mr. Wei LIU


Abstract:

Language models increasingly solve problems by producing intermediate 
reasoning and by interacting with external tools and environments. 
Reinforcement learning provides a common way to improve these behaviors from 
feedback, but the meaning and reliability of that feedback change 
substantially across tasks. This survey reviews the transition from reasoning 
on static inputs to acting in environments through four connected dimensions: 
reward construction, credit assignment, interaction horizon, and computational 
cost. We begin with reinforcement learning in verifiable environments, 
training dynamics, and stable policy optimization under training-inference 
mismatch. Within reasoning, we cover efficient computation allocation and 
multimodal extensions. We then discuss tool use, deep research, and general 
tool-using agents before turning to coding tasks, agent harnesses, and 
learning from execution. Frontend applications, scientific workflows, and 
kernel optimization illustrate advanced coding settings with different 
feedback requirements. Finally, we identify open questions about 
generalization, long-horizon learning, and controlled self-improvement. The 
emphasis is on distinguishing empirical evidence from stronger claims about 
capability and on separating improvements due to policy learning, data, 
test-time scaling, and agent interfaces.


Date:                   Friday, 16 October 2026

Time:                   10:00am - 12:00noon

Venue:                  Room 5504
                        Lifts 25/26

Committee Members:      Dr. Junxian He (Supervisor)
                        Dr. May Fung (Chairperson)
                        Dr. Ling Pan