ScopeEdit enhances multimodal LLM editing by controlling knowledge propagation

Researchers have introduced ScopeEdit, a novel method for online multimodal knowledge editing in large language models (MLLMs). This approach aims to control the semantic scope of each edit, ensuring that corrections transfer to relevant cross-modal variants without negatively impacting unrelated inputs. ScopeEdit decomposes updates into modality-local and shared generalization branches, utilizing orthogonal low-rank spaces and Sherman-Morrison recursions for efficient, bounded overhead. Experiments demonstrate ScopeEdit's effectiveness in improving the balance between in-scope transfer and out-of-scope locality across various benchmarks and MLLM architectures. AI

IMPACT Improves the precision and control of knowledge updates in multimodal LLMs, potentially leading to more robust and reliable AI systems.

RANK_REASON The item is a research paper detailing a new method for multimodal LLM editing. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• multimodal large language model
• ScienceCast
• ScopeEdit
• Sherman--Morrison
• VLKEB

• paper
• model release

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →