Incremental Context and User Adaptation for Large Language Models

PhD Thesis Proposal Defence

Title: "Incremental Context and User Adaptation for Large Language Models"

by

Miss Yeongseo JUNG


Abstract:

Large language models are increasingly deployed as persistent interactive 
systems. Sustained interaction introduces a different generation problem. 
Information accumulated over time is dynamic: its relevance depends on the 
current context, earlier information may need to be reconsidered, and 
user-specific patterns emerge as evidence accumulates. Therefore, effective 
interaction requires models to continually determine how previously available 
and newly observed information should shape subsequent behavior. However, 
repeatedly processing the complete interaction history or continually adapting 
a large model incurs substantial computational overhead.

This thesis investigates these challenges through three complementary forms of 
adaptation over task knowledge, interaction memory, and user state. 
Contextualized Knowledge Distillation integrates recommendation and dialogue 
knowledge through context-specific gating, enabling different sources of task 
knowledge to guide generation according to the current conversational context. 
Context-Driven Incremental Compression maintains compact and revisable 
dialogue memory, allowing previously processed information to be retrieved, 
updated, and reused without repeatedly encoding the complete history. Bayesian 
User-State Inference formulates personalization as sequential inference over a 
latent user state, incrementally updating its belief from sparse and evolving 
observations while keeping the language model fixed.

Together, these studies advance interactive language models along three 
complementary dimensions: context-sensitive coordination of heterogeneous 
capabilities, persistent and revisable long-term memory, and online inference 
of evolving user state. They move beyond static generation from a fixed prompt 
toward systems that can adapt their behavior as conversational context, 
accumulated information, and user evidence change over time. By supporting 
these forms of adaptation without continual backbone optimization or 
computation that grows with the complete interaction history, this thesis 
advances toward scalable language models for sustained interaction.


Date:                   Friday, 18 September 2026

Time:                   2:00pm - 4:00pm

Venue:                  Room 5501
                        Lift 25/26

Committee Members:      Prof. Lei Chen (Supervisor)
                        Prof. Qiong Luo (Chairperson)
                        Dr. May Fung