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

Abstract

Unified multimodal models (UMMs) have established themselves as a leading framework for general-purpose multimodal intelligence. However, as these systems are integrated into real-world scenarios, the ability to efficiently update their internal knowledge bases becomes a paramount concern. Although knowledge editing techniques have become well-developed for text-only architectures, it remains uncertain whether successful textual modifications also extend to image generation capabilities within UMMs. To address this uncertainty, we present UniKE, the inaugural benchmark designed for cross-modality knowledge editing in UMMs. This dataset includes 2,971 edit subjects covering both attribute and relational changes.

Through visual verification utilizing VQA, our analysis uncovers a significant modality gap. While efficacy on the textual side can achieve approximately 92%, the best overall VQA accuracy observed during direct image generation stands at a mere 18.5%. To bridge this divide, we introduce Reasoning-augmented Parameter Editing, a method that explicitly activates edited knowledge prior to generation. This approach enhances overall VQA accuracy across all tested model-editor combinations, yielding improvements of up to 18.6 percentage points.

Mechanistic investigations indicate that this performance disparity stems from a partial misalignment between the edited textual representations and the conditioning pathways responsible for visual generation. Specifically, modifications that are adequate for influencing text outputs may be insufficiently strong or incorrectly aligned to effectively direct image synthesis. These results demonstrate that textual knowledge edits do not ensure reliable transfer across modalities, thereby highlighting the need for editing methods that are aware of modality-specific nuances. Our code and data are available at https://github.com/gxx27/UniKE.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC