AdUE: Improving uncertainty estimation head for LoRA adapters in LLMs
This addresses uncertainty estimation for classification tasks in language models using parameter-efficient fine-tuning, which is an incremental improvement over existing methods.
The paper tackles the problem of uncertainty estimation in parameter-efficient fine-tuning of language models for classification tasks by introducing AdUE, a post-hoc method that improves softmax-based estimates through differentiable approximation and regularization. The method consistently outperforms established baselines across five NLP datasets and four language models, producing better-calibrated confidence with a lightweight approach.
Uncertainty estimation remains a critical challenge in adapting pre-trained language models to classification tasks, particularly under parameter-efficient fine-tuning approaches such as adapters. We introduce AdUE1, an efficient post-hoc uncertainty estimation (UE) method, to enhance softmax-based estimates. Our approach (1) uses a differentiable approximation of the maximum function and (2) applies additional regularization through L2-SP, anchoring the fine-tuned head weights and regularizing the model. Evaluations on five NLP classification datasets across four language models (RoBERTa, ELECTRA, LLaMA-2, Qwen) demonstrate that our method consistently outperforms established baselines such as Mahalanobis distance and softmax response. Our approach is lightweight (no base-model changes) and produces better-calibrated confidence.