Mitigating Overconfidence in Large Language Models: A Behavioral Lens on Confidence Estimation and Calibration
Dec 1, 2024·
,,,,,·
1 min read
Bingbing Wen
Chenjun Xu
Bin Han
Robert Wolfe
Lucy Lu Wang
Bill Howe
Abstract
We take a behavioral perspective on confidence estimation and calibration in large language models, proposing methods to mitigate overconfidence by aligning model confidence with reliability.
Type
Publication
NeurIPS 2024 Workshop on Behavioral Machine Learning
We examine overconfidence in large language models through a behavioral lens and propose approaches to improve calibration so that stated confidence better reflects actual reliability.
