Yizhou Chi

PhD thesis: Building Trustworthy LLMs: Tackling Misinformation, Dishonesty, and Contextual Inconsistency

Research interests:

1. Trustworthy AI
2. Multi-agent orchestration
3. Automated fact checking
4. AI Safety and self-regulation

My research seeks to address three critical challenges with Large Language Models (LLMs): misinformation through hallucinations, dishonesty through overconfidence, and groundlessness through inconsistent context maintenance. I aim to develop robust methodologies to enhance LLMs' factual reliability, transparency, and contextual consistency while preserving their flexibility and creativity. My approach centres on three interconnected research directions: (1) Developing advanced fact-checking systems that enable LLMs to self-monitor, verify claims, and recognize hallucinations; (2) Implementing self-regulation mechanisms including self-revision and restraint capabilities that allow LLMs to improve outputs and express appropriate uncertainty; and (3) Creating clarification-seeking frameworks that help LLMs recognize ambiguity and maintain consistency in extended interactions. These contributions will support the responsible and trustworthy deployment of LLMs across various industries while mitigating the risks of false information proliferation.

Who or what inspired you to pursue your research interests?

Recent advancements in Natural Language Processing have demonstrated the impressive capabilities of Large Language Models (LLMs). However, one of my primary concerns is that as LLMs become indispensable in various industries, we may be overwhelmed by an increase in false or deliberately manipulated information. This flood of misinformation, combined with the challenge of interacting with inconsistent agents, could significantly hinder progress across multiple fields. My firsthand experience developing AI systems revealed these fundamental limitations, inspiring my commitment to creating mechanisms that enable AI to verify information, express appropriate uncertainty, and maintain coherence across extended interactions—ultimately building more trustworthy AI systems that can be responsibly integrated into society.