Knowledge-Sensitive Dynamic Module Editing: Precise Knowledge Revision for Multimodal Large Language Models

Keyu Ren, Zilei Wang

Keywords: Multimodal Large Language Models, Knowledge Editing

Abstract: Multimodal Large Language Models (MLLMs) struggle with efficient knowledge updates because their internal representations distribute information across lengthy and heterogeneous visual-textual sequences. This distribution makes traditional "locate-then-edit" methods, despite being highly effective in text-only models, largely ineffective for MLLMs. The resulting challenges include inaccurate localization of knowledge, poor generalization of edits, and unintended damage to unrelated knowledge. To bridge this gap, we introduce KDKE, a novel Knowledge-sensitive Dynamic multimodal Knowledge Editing framework tailored for MLLMs. KDKE introduces an Integrated Module Contribution Score to precisely quantify the impact of different modules on specific knowledge outputs. This enables a dynamic module selection mechanism that identifies critical parameters for each edit instance adaptively. We further develop a constrained adaptive editing algorithm, which injects LoRA parameters into selected modules and optimizes them under multi-objective constraints to ensure reliable editing, robust generalization, and strict locality. Extensive experiments on multiple model architectures and benchmarks demonstrate that KDKE superior editing accuracy and consistently strong overall performance, providing an effective and reliable solution for knowledge editing in multimodal settings.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 909