Data Science Talk Series 

Title: Abstention in Large Language Models: Towards Enhanced Reliability and Alignment 

Abstract: Large language models (LLMs) encounter significant challenges when responding under uncertainty, making reliable abstention—where models opt not to answer uncertain queries—an important capability. In this talk, I present recent research focused on developing benchmarks specifically designed for evaluating abstention behaviors in LLMs. Additionally, I discuss how confidence estimation methods inspired by human psychological insights on overconfidence can improve the calibration of model responses. Finally, I introduce a comprehensive survey that integrates human values and LLM's confidence into a structured framework aimed at advancing LLM abstention strategies. These efforts collectively contribute to creating safer, more reliable, and human-aligned AI systems.