Bayesian Sparse Low-Rank Adaptation for Large Language Model Uncertainty Estimation