Grounding LLM Agents in Knowledge, Context, and Action

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


PhD Thesis Defence


Title: "Grounding LLM Agents in Knowledge, Context, and Action"

By

Mr. Jiayang CHENG


Abstract:

Large language models (LLMs) are increasingly deployed as autonomous agents, 
yet they remain insufficiently grounded in the information they rely on. This 
thesis studies LLM agent grounding along three dimensions. For knowledge, we 
develop methods for detecting and resolving conflicts among retrieved 
evidence, finding that even strong LLMs frequently favor one piece of 
conflicting evidence without justification. We further introduce an evaluation 
framework for verifying claims that require multiple interdependent pieces of 
evidence, revealing that models struggle with partially supported claims and 
tend to compensate for missing information using internal knowledge. For 
context, we build an interactive benchmark that evaluates how well agents 
maintain memory over extended conversations through on-policy interaction, 
uncovering significant memory limitations across both standalone LLMs and 
LLM-powered agents. For action, we train agents to orchestrate multi-step API 
calls through reinforcement learning with a graduated reward design that 
decomposes correctness into atomic validity and orchestration consistency, 
improving performance on complex tool-use tasks. Collectively, these 
contributions provide evaluation tools and training methods that help LLM 
agents operate more reliably in real-world settings.


Date:                   Thursday, 21 May 2026

Time:                   10:00am - 12:00noon

Venue:                  Room 2128A
                        Lift 19

Chairman:               Prof. Vincent Kin Nang LAU (ECE)

Committee Members:      Dr. Yangqiu SONG (Supervisor)
                        Prof. Raymond WONG
                        Dr. Dan XU
                        Prof. Bert SHI (ECE)
                        Dr. Xiao-Ming WU (PolyU)