Please register for the lecture here.

Abstract: 

Large Language Models (LLMs) have shown remarkable capabilities across various tasks, but their deployment in high-stake domains relies on sound evaluation. In this talk, I will present work on evaluating one aspect of the robustness of LLM models – their consistent and coherent behavior across multiple rounds of user interaction. This work introduces a comprehensive framework for evaluating and improving LLM response consistency, making three key contributions. First, we introduce Position-Weighted Consistency (PWC), a metric designed to capture both the importance of early-stage stability and recovery patterns in multi-turn interactions. Second, we present MT-Consistency, a carefully curated benchmark dataset spanning diverse domains and difficulty levels, specifically designed to evaluate LLM consistency under various challenging follow-up scenarios. Third, we introduce Confidence-Aware Response Generation (CARG), a framework that significantly improves response stability by explicitly integrating internal model confidence scores during the generation process. Focusing on the case when models switch from correct answers to incorrect answers in the face of adversarial prompting, we statistically analyze the behavior of the models using ideas drawn from survival analysis to develop actionable insights for societally consequential applications. More information is available in a forthcoming ACL paper.

This research was conducted in collaboration with Yubo Li, Yidi Miao, Xueying Ding, and Rema Padman.