New Method Improves Knowledge Editing in Multimodal LLMs

Multimodal Large Language Models (MLLMs) face a challenge where knowledge updates applied using combined text-image inputs often fail to persist when the model is queried with single-modality inputs. This issue, termed "editing decoupling failure," means that information learned in a multimodal context isn't consistently accessible through unimodal pathways. An in-depth analysis revealed that MLLMs store entity knowledge in separate, modality-specific pathways rather than a unified representation. Consequently, updates optimized for multimodal queries do not effectively propagate to these individual unimodal circuits. To address this, a new framework called DECODE has been introduced. DECODE explicitly works to disentangle and pinpoint neuron groups responsible for specific modalities, allowing for more targeted and effective knowledge updates. Experiments show that DECODE successfully ensures consistent knowledge updates regardless of whether the model is triggered by multimodal or unimodal inputs, thereby resolving the decoupling problem.