DALorRA applies Bayesian sparse LoRA to improve LLM uncertainty estimation

The authors propose Data-Adaptive Lower-Rank Adaptation (DALorRA), a variational Bayesian sparse framework designed to address the overconfidence issue in large language models during task-specific fine-tuning. This method shifts uncertainty quantification from the dense parameter space to the lightweight rank level of low-rank adaptation (LoRA). By imposing stochastic masking on rank dimensions, DALorRA enables Bayesian regularization of model capacity during training and provides ensemble-like calibration during inference.

Extensive experiments demonstrate that DALorRA achieves excellent calibration for LLMs without compromising their reasoning accuracy.