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

This paper asks whether text-only knowledge edits in unified multimodal models actually transfer to image generation, and the answer is mostly no. The authors introduce U NI KE, the first benchmark for cross-modality knowledge editing in UMMs, built with 2,971 edit subjects and 5,535 evaluation instances spanning attribute and relation edits, then test three representative models—Ovis-U1, BLIP3o-4B, and OmniGen2—using parameter editors MEMIT, PMET, and AlphaEdit. The central finding is a sharp modality gap: text-side edit efficacy reaches about 92% in some settings, but the best direct image-generation VQA accuracy is only 18.5%, showing that successful factual rewriting in text does not reliably change generated images. A reasoning-augmented protocol that first forces the model to verbalize the edited fact improves every model-editor pair, with gains as large as 18.6 percentage points, but it still leaves a large gap. Mechanistically, the paper traces the failure to a conditioning-pathway bottleneck: for Ovis-U1, edit perturbations are heavily attenuated by a frozen Linear(4096, 1536) projection, retaining only about 34.82% of perturbation energy in the top 1,536 directions, whereas reasoning expands the effective conditioning shift by 8.56× and yields the largest visual gains. Overall, the study shows that cross-modal knowledge editing requires more than changing text outputs; the edit must survive the language-to-vision interface and align with the image generator’s conditioning pathway.