Attribution Analysis Meets Model Editing: Advancing Knowledge Correction in Vision Language Models with VisEdit

TL;DR

This paper introduces VisEdit, a novel method for editing visual representations in vision-language models to correct knowledge errors, demonstrating significant improvements over existing approaches through attribution analysis and benchmark evaluations.

Contribution

The paper pioneers the application of model editing to VLLMs by developing VisEdit, which edits intermediate visual representations to correct knowledge without retraining.

Findings

VisEdit outperforms state-of-the-art baselines in VLLM editing tasks.

Visual representations in mid-to-later layers significantly influence predictions.

Attribution methods effectively identify regions relevant to knowledge edits.

Abstract

Model editing aims to correct outdated or erroneous knowledge in large models without costly retraining. Recent research discovered that the mid-layer representation of the subject's final token in a prompt has a strong influence on factual predictions, and developed Large Language Model (LLM) editing techniques based on this observation. However, for Vision-LLMs (VLLMs), how visual representations impact the predictions from a decoder-only language model remains largely unexplored. To the best of our knowledge, model editing for VLLMs has not been extensively studied in the literature. In this work, we employ the contribution allocation and noise perturbation methods to measure the contributions of visual representations for token predictions. Our attribution analysis shows that visual representations in mid-to-later layers that are highly relevant to the prompt contribute significantly to predictions. Based on these insights, we propose VisEdit, a novel model editor for VLLMs that effectively corrects knowledge by editing intermediate visual representations in regions important to the edit prompt. We evaluated VisEdit using multiple VLLM backbones and public VLLM editing benchmark datasets. The results show the superiority of VisEdit over the strong baselines adapted from existing state-of-the-art editors for LLMs.