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

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.

AI models that can both understand and create images store vast amounts of factual knowledge — like knowing that apples are typically red. But what happens when we need to correct or update a fact inside such a model? We found that even when a model successfully learns a new fact in text (say, "this apple is blue"), it often fails to reflect that change when generating images. The apple still comes out red. We traced this failure to an architectural bottleneck: in some models, a narrow internal bridge connects the language system to the image generator. This bridge filters out most of the edited signal before it ever reaches the part that draws pictures — like whispering a correction through a thick wall. To work around this, we developed a simple strategy: before generating an image, we first ask the model to reason out loud about what it knows. This reasoning text then feeds back into the image generator through a much wider channel — the same pathway the model already uses to turn ordinary descriptions into pictures. For models with the worst bottleneck, this "think before you draw" approach nearly tripled image accuracy. Our work reveals a fundamental tension in multimodal AI: systems that separate language from vision for efficiency may inadvertently block knowledge from flowing between the two. Understanding where and why this happens is a first step toward building AI systems whose knowledge stays consistent across everything they say and show.