Understanding the Sources of Uncertainty for Large Language and Multimodal Models
Keywords: LLM, uncertainty decomposition, hallucination, VLM
TL;DR: Decompose the predictive uncertainty of a pretrained model into separate components, identify the calibrated parts that can be used for detecting hallucinations and enhancing performance.
Abstract: Understanding and quantifying uncertainty in large model predictions is critical for their safe and trustworthy deployment. However, existing methods that estimate the overall prediction uncertainty often fail due to miscalibration like model overconfidence. Uncertainty decomposition provides a way to focus on some specific parts in total uncertainty, removing those unrelated components. Traditional uncertainty decomposition into epistemic (model-related) and aleatoric (data-related) components is insufficient for current model usage, as additional factors like prompt phrasing and context significantly influence the model’s predictions and add the source of uncertainty. We introduce a unified uncertainty decomposition framework that systematically separates uncertainty contributed by various factors such as prompting, context, and preprocessing of multimodal inputs. By quantifying each component’s uncertainty, our approach identifies which uncertainty terms are well-correlated with the model’s hallucination rates, thereby enhancing hallucination detection and model improvement. We validate our framework through applications in visual question answering and math reasoning, demonstrating that effective uncertainty components can serve as metrics for hallucination detection and improve model performance through self-training. Grounded in information theory and highly extensible, our framework provides a novel perspective on uncertainty decomposition in large language and multimodal models, offering valuable insights for future research.
Submission Number: 30
Loading