HMDR Method for Fake News Detection with Missing Modalities
October 28, 2026
The Head-wise Modality Decoupling and Regulation (HMDR) method improves multimodal fake news detection by separating modality representations through soft head grouping. This approach compensates for imbalanced attention when images or text are missing, reducing modality bias in multimodal LLMs.
HOW THIS AFFECTS YOU
●
researcherYou can implement this decoupling method to improve model robustness in multimodal datasets with missing inputs.