Correct When Paired, Wrong When Split: Decoupling and Editing Modality-Specific Neurons in MLLMs

The paper introduces DECODE, a method designed to address the issue of editing decoupling failure in Multimodal Large Language Models (MLLMs), where knowledge updates from multimodal inputs do not propagate effectively to unimodal inputs. The authors provide empirical evidence showing that knowledge is distributed across modality-specific pathways, leading to inconsistent updates. DECODE explicitly disentangles these neuron groups, enabling targeted knowledge updates that mitigate the failure, which is crucial for practitioners aiming to enhance the reliability of MLLMs in diverse input scenarios.