DeLo: Dual Decomposed Low-Rank Experts Collaboration for Continual Missing Modality Learning

Researchers propose DeLo, a new framework using dual-decomposed low-rank expert architecture to help Large Multimodal Models adapt to real-world scenarios with incomplete data. The system addresses continual missing modality learning by preventing interference between different data types and tasks through specialized routing and memory mechanisms.

• →DeLo introduces the first dual-decomposed low-rank expert architecture specifically designed for continual missing modality learning in large multimodal models.
• →The framework resolves modality interference through decomposed LoRA experts that dynamically compose update matrices from separate modality-specific factor pools.
• →A task-partitioned structure prevents catastrophic forgetting while Cross-Modal Guided Routing handles incomplete data scenarios.
• →Experimental results show significant performance improvements over existing state-of-the-art approaches on established benchmarks.
• →The research highlights the importance of principled architectural design for real-world multimodal AI applications.