Event Details

PhD Thesis Proposal - Justin Cho

Wed, Jul 16, 2025
1:30 PM - 3:30 PM
Location: EEB 112
 

Title: Improving Human-AI Interactions via Context Synchronization
 
 
Date and Time: Wednesday, July 16th | 1:30p - 3:30p
 
 
Location: EEB 132
 
 
Committee Members: Jonathan May, Maja Mataric, Shrikanth Narayanan, Robin Jia, Swabha Swayamdipta 
 
 
Abstract: A shared understanding of context - the circumstances that form the setting for an event, statement, or idea, and in terms of which it can be fully understood and assessed - is imperative for human-AI interactions to be successful. For example, for a language model (LM) assistant to effectively help a student aspiring to learn about the mechanisms of modern LMs, it must be aware of the student's higher-level goal, their background knowledge, and the mode of interaction. The continuous refinement and expansion of context that LMs harness have formed the cornerstone of their impressive feats of surpassing human performance on various benchmarks. However, when put to the test with real-world tasks, LMs  remain insufficient because their contextual understanding is not synchronized with that of the human user in various dimensions.
 
 
In this thesis, I present a new paradigm, context synchronization, that discovers and bridges the remaining gaps in the context shared by the human and assisting AI and thus facilitate the knowledge transfer between them for safe and successful human-AI outcomes. In Part I, I introduce benchmarks that measures the reliability of AI models for non-text modes of communication, speech and nonverbal communication, and thus provide insights for their applications to use cases that require contextual understanding that extends beyond a textual interface. In Part II, I develop literature-grounded computational methods that contextualizes language models to higher-order objectives beyond transactional interactions. In Part III, I present data-efficient computational methods for contextualizing language modes to individual users that process user-specific data to generate personalized text using language models.
 
 
The work in this thesis has broad implications for versatile and safe AI assistants that deliver on their promise of accessible and democratized utility. It demonstrates that by evaluating and adapting AI systems to maximize shared context between the human and AI counterpart can help tackle complex and nuanced real-world tasks. By discussing the challenges of methods, evaluations, and applications, this thesis advances a human-centric vision of AI grounded in comprehensive context synchronization.