ScopeEdit Controls Multimodal LLM Knowledge Edits, Preventing Leakage

Online knowledge editing for multimodal large language models (MLLMs) involves continually injecting visual-textual corrections with minimal disruption to unrelated behaviors. Existing editing methods often focus on reliability and stability but frequently fail to control the precise semantic scope of each edit, leading to unintended side effects. New research identifies a "scope gap" where successful instance-level edits don't guarantee transfer to relevant cross-modal variants and can leak to unrelated inputs. To address this, the paper formulates "Edit-Scoped Generalization," reframing MLLM editing to control the propagation boundary of each update. The proposed solution, ScopeEdit, is a scope-aware online editor that decomposes updates into a modality-local absorption branch and an evidence-gated shared generalization branch. The local branch handles stable edit absorption, while the shared branch enables cross-modal propagation only when visual and textual evidence are sufficiently aligned. Both branches use scope-separated write geometries and maintain preconditioners, ensuring constant per-edit overhead. Extensive experiments confirm ScopeEdit's superior trade-off between in-scope cross-modal transfer and out-of-scope locality, while preserving edit reliability and efficiency.