DECODE fixes MLLM editing failure when inputs split across modalities

A team led by Tingchao Fu has uncovered a significant limitation in current knowledge editing techniques for Multimodal Large Language Models (MLLMs). While editing can successfully update entity-related knowledge when the model is triggered by paired text-image queries, those same edits often fail when the inputs are split into unimodal queries (text-only or image-only). The root cause: MLLMs do not store entity knowledge as a unified representation. Instead, knowledge is distributed across disentangled, modality-specific neural pathways. Edits biased toward multimodal inputs simply do not propagate to the unimodal circuits, leaving outdated facts intact in those pathways.

To address this, the authors propose DECODE, a method that explicitly decouples and localizes modality-specific neuron groups for targeted knowledge updates. By identifying which neurons are responsible for text-based vs. image-based knowledge, DECODE ensures that edits apply consistently regardless of input modality. Extensive experiments demonstrate that DECODE successfully mitigates the editing decoupling failure, achieving robust knowledge updates under different modality triggers. The 18-page paper, with 11 figures, provides empirical evidence that decoupling modality-specific neurons is essential for reliable MLLM updates.