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

In three linesResearchers identify a critical issue in knowledge editing for MLLMs: updates work with multimodal inputs (text+image) but fail with unimodal inputs alone. They propose DECODE, a method that localizes and decouples modality-specific neurons to propagate edits consistently across all input types.

Summary generated by Claude — human-verified