Title: Toward Reliable LLM Agents in Healthcare: A Synthesis of Self-Improvement Mechanisms

Presented by Qijia Liu, Computing PhD student, Data Science emphasis

Abstract

The tendency to use large language models (LLMs) in clinical and public-health fields is becoming increasingly prevalent, especially in supporting medical decision-making. However, the presence of overconfident but incorrect answers or hallucinations, represents an existing hidden risk that could lead to serious consequences in healthcare. This paper investigates three self-improvement mechanisms that aim to mitigate such problems in agentic LLM systems without tuning model parameters: I begin by reviewing Reflexion (internal self-critique) as an example of internal verification; next, I examine AgentsCourt (role-play debating) as an example of adversarial verification; and finally, I detail ProTeGi as a framework for external optimization, which optimizes prompts using textual gradient descent and beam search at the prompt level. After a brief introduction of my artifact, I discuss the limitations of these three mechanisms and how they could collectively contribute to safer healthcare applications.

Advisor: Dr. Edoardo Serra

Committee Members: Dr. Antonino Rullo, Dr. Bogdan Dit

External Examiner: Dr. Max Taylor