Before Fusion, Ask What to Keep: Contextual Calibration of Multimodal Signals

In three linesPre-fusion calibration method for multimodal signals. A compact module compares each modality (language, sound, vision) against others, extracts cross-source support and conflict cues, and modulates representations before merging. Tested on 5 benchmarks (sentiment, action recognition, audio-visual event detection, emotion classification) with consistent improvements.

Summary generated by Claude — human-verified