Researchers propose ScopeEdit, a scope-aware online editor that reframes multimodal large language model (MLLM) editing by controlling the propagation boundary of each update rather than merely correcting instances. The method decomposes updates into a modality-local absorption branch and an evidence-gated shared generalization branch to manage cross-modal transfer.
- ScopeEdit uses orthogonal low-rank spaces for scope-separated write geometries and Sherman--Morrison recursions for preconditioners, yielding constant per-edit overhead.
- The shared branch enables cross-modal propagation only when visual and textual evidence are sufficiently aligned.
- Experiments across diverse benchmarks and long-horizon edit streams show improved trade-offs between in-scope transfer and out-of-scope locality while preserving reliability and stability.
The approach addresses the scope gap in existing editors by preventing leakage to unrelated inputs and ensuring edits do not generalize beyond their intended semantic boundary.