Chain of Modality: From Static Fusion to Dynamic Orchestration in Omni-MLLMs
Omni-modal Large Language Models struggle with static fusion methods that cause performance fragility due to structural flaws like positional bias and alignment traps. The Chain-of-Modality (CoM) framework solves this by dynamically orchestrating how different modalities interact based on the task, switching between parallel, sequential, and interleaved pathways to achieve superior robustness and efficiency.