LPEdit: Locality-Preserving Knowledge Editing for MultiModal Large Language Models

ACL ARR 2025 May Submission6082 Authors

Abstract: Large Language Models (LLMs) and Vision Large Language Models (VLLMs) demonstrate impressive abilities in comprehending natural language and interpreting visual information, but they can also preserve outdated or incorrect information in both forms. Existing knowledge editing methods can efficiently update erroneous text information in LLMs, avoiding the need for full retraining. The locality in multimodal knowledge editing refers to editing that should affect only the targeted outputs while preserving the model’s behavior on unrelated inputs in both textual and visual modalities. Existing methods often overlook this principle and do not explicitly design to preserve the consistency of responses on unrelated information. Here, we propose LPEdit, a novel method that leverages the null space projection on key layers to focus the editing on conveyed visual information without influencing unrelated knowledge. Experiments show that our method achieves strong performance across different models and datasets. Moreover, our work advances the understanding and development of locality in multimodal knowledge editing.

Paper Type: Short

Research Area: Interpretability and Analysis of Models for NLP

Research Area Keywords: model editing, multimodal applications

Contribution Types: NLP engineering experiment

Languages Studied: English

Keywords: model editing, multimodal applications

Submission Number: 6082