Geometry-Calibrated Conformal Abstention for Language Models

Rui Xu, Yi Chen, Sihong Xie +1

When language models lack relevant knowledge for a given query, they frequently generate plausible responses that can be hallucinations, rather than admitting being agnostic about…