New Method Addresses Knowledge Editing Inconsistencies in Multimodal LLMs

2026-06-17

Researchers have identified a knowledge editing decoupling failure in Multimodal Large Language Models (MLLMs), where updates made with paired text-image inputs do not persist when inputs are separated. A new method, DECODE, aims to resolve this by localizing and disentangling modality-specific neurons for targeted knowledge updates.

Source: arXiv · cs.LG

Reported by VERA Newswire.