GAUGE Enhances Multimodal Classification with Incomplete Data.

Multimodal classification often assumes all modalities are available, but real-world data frequently suffers from incompleteness. Existing methods for handling missing modalities, such as imputation or dynamic fusion, typically operate at a coarse modality level, failing to selectively retain reliable components while suppressing misleading ones within a single recovered modality. GAUGE (Granularity-Adaptive Counterfactual Gating of Evidence) addresses this by providing a lightweight framework that operates at a fine-grained evidence unit level. It first imputes missing modalities and then encodes all inputs uniformly. Instead of explicit intervention, GAUGE uses prediction-aware Taylor evidence scores, derived from a single forward-backward pass, to quantify the counterfactual effect of each unit. These scores are then mapped to continuous gates, which apply additive attention-logit biases for unit-wise evidence modulation, crucially without modifying the backbone architecture. Experiments across six benchmarks show GAUGE outperforming strong baselines in diverse incomplete-input scenarios, establishing it as a principled and scalable solution for fine-grained evidence control.