Do Text Edits Generalize to Visual Generation? Benchmarking Cross-Modal Knowledge Editing in UMMs

2026-05-29

AI Summary

This work addresses the significant cross-modal knowledge transfer gap in existing unified multimodal models, where textual knowledge editing fails to generalize effectively to image generation. To tackle this challenge, the authors introduce UniKE—the first benchmark for cross-modal knowledge editing, comprising 2,971 edit subjects—and propose a reasoning-enhanced parameter editing method that explicitly activates edited knowledge prior to generation to improve visual consistency. Evaluation via vision-language question answering (VQA) reveals that while text-only edits achieve a high success rate of 92%, their direct application to image generation yields a VQA accuracy of only 18.5%. The proposed approach improves this metric by up to 18.6 percentage points, substantially enhancing cross-modal editing performance and advancing modality-aware knowledge editing.

Abstract

Unified multimodal models (UMMs) have emerged as a promising paradigm for general-purpose multimodal intelligence. As they are deployed in real-world applications, effectively updating internal knowledge becomes critical. While knowledge editing has matured for text-only models, it remains unclear whether edits that successfully modify textual outputs also transfer to image generation in UMMs. To study this question, we introduce UniKE, the first benchmark for cross-modality knowledge editing in UMMs, comprising 2,971 edit subjects spanning attribute and relation edits. Using VQA-based visual verification, we reveal a striking modality gap: text-side efficacy can reach approximately 92%, whereas the best overall VQA accuracy under direct image generation is only 18.5%. We further propose Reasoning-augmented Parameter Editing, which explicitly activates edited knowledge before generation and improves overall VQA accuracy for all evaluated model-editor pairs, with gains up to 18.6 percentage points. Mechanistic analysis shows that this gap is associated with partial alignment between edited textual representations and the conditioning pathways for visual generation, where edits sufficient for text outputs may remain too weak or misaligned to steer image synthesis. These findings show that textual knowledge edits do not guarantee reliable cross-modality transfer and motivate modality-aware editing methods. Our code and data are available at https://github.com/gxx27/UniKE.

Problem

• knowledge editing
• cross-modal generalization
• unified multimodal models
• visual generation
• modality gap

Innovation

• cross-modal knowledge editing
• unified multimodal models
• knowledge editing
• visual generation
• modality gap