Multimodal Knowledge Edit-Scoped Generalization for Online Recursive MLLM Editing

The rapid development of multimodal large language models (MLLMs) has opened new horizons in AI, enabling machines to understand and generate both visual and textual information seamlessly. However, maintaining up-to-date knowledge in these models remains a challenge, especially when updates need to be performed online without retraining from scratch. Existing knowledge editing methods primarily focus on ensuring the reliability and stability of edits but often overlook the crucial aspect of controlling the scope of knowledge propagation.

This oversight can lead to unintended consequences, such as knowledge leakage into unrelated inputs or insufficient transfer to semantically similar variants. To address this, the authors propose ScopeEdit, a novel framework that introduces explicit scope control into online multimodal knowledge editing. The core idea is to decompose each update into two components: a modality-local absorption branch that guarantees stable, reliable editing, and a shared generalization branch that propagates knowledge only when supported by sufficient multimodal evidence. This is achieved through a scope-aware mechanism that employs orthogonal low-rank geometries and evidence gating, ensuring that updates are activated only under semantic alignment.

The technical innovation lies in the use of dual low-rank spaces for scope separation, combined with recursive preconditioning to maintain constant complexity per update. This design allows the model to balance the competing demands of transferability and locality, significantly improving the model’s ability to generalize within the intended scope while preventing leakage outside it. Extensive experiments across benchmarks, real-world scenarios, and diverse architectures demonstrate that ScopeEdit outperforms existing methods in cross-modal transfer success, stability over long sequences, and locality preservation.

The broader impact of this work is substantial, providing a scalable, efficient, and controllable solution for continuous knowledge updates in complex multimodal systems. It paves the way for safer, more reliable AI capable of ongoing learning in dynamic environments. Nonetheless, challenges remain in adaptive threshold tuning and handling ambiguous evidence, which future research aims to address, further advancing the field of controllable AI systems.