GAUGE: A Counterfactual Gating Framework for Incomplete Multimodal Classification

A new research paper on arXiv (ID: 2608.05608) introduces GAUGE, a lightweight counterfactual gating framework designed to handle incomplete multimodal classification. The framework addresses the common real-world problem where some modalities are missing during classification tasks. Traditional methods either impute missing data or dynamically fuse available modalities, but they operate at a coarse modality level, failing to distinguish reliable components from misleading ones within a recovered modality. GAUGE overcomes this by first imputing missing modalities with a frozen imputer, then encoding both observed and recovered inputs as fine-grained evidence units. Instead of intervening on each unit explicitly, it computes prediction-aware Taylor evidence scores to estimate the counterfactual effect of replacing each unit with a reference representation, all in a single forward-backward pass. These scores are then mapped to continuous gates that control the influence of each evidence unit, allowing the model to retain reliable information while suppressing misleading components. The paper is categorized as a cross-type announcement and is available at https://arxiv.org/abs/2608.05608.