Research map / Vision-language and multimodal / Cross-modal learning
Do Text Edits Generalize to Visual Generation? Benchmarking Cross-Modal Knowledge Editing in UMMs
Xin Gao, Cheng Yang, Chufan Shi, Taylor Berg-Kirkpatrick
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add reasoning-augmented generation (edited model verbalizes edited fact first) (ablation) | PMET on Ovis-U1 · UniKE Overall | VQA accuracy (%): 9.71 -> 28.32 | Improved |
| add reasoning-augmented generation (ablation) | MEMIT on Ovis-U1 · UniKE Overall | VQA accuracy (%): 8.70 -> 24.41 | Improved |
| add reasoning-augmented generation (ablation) | AlphaEdit on Ovis-U1 · UniKE Overall | VQA accuracy (%): 7.42 -> 20.67 | Improved |
| add reasoning-augmented generation (ablation) | PMET on BLIP3o-4B · UniKE Overall | VQA accuracy (%): 18.51 -> 19.29 | Improved |
| add reasoning-augmented generation (ablation) | AlphaEdit on BLIP3o-4B · UniKE Overall | VQA accuracy (%): 16.12 -> 17.33 | Improved |
| add reasoning-augmented generation (ablation) | AlphaEdit on OmniGen2 · UniKE Overall | VQA accuracy (%): 11.50 -> 17.90 | Improved |
| add reasoning-augmented generation (ablation) | MEMIT on OmniGen2 · UniKE Overall | VQA accuracy (%): 8.73 -> 12.17 | Improved |
| use PMET instead of MEMIT | Ovis-U1 direct · UniKE Overall | text-side efficacy (%): 59.84 -> 72.18 | Improved |
| use softened null-space AlphaEdit instead of PMET | BLIP3o-4B reasoning-augmented · UniKE Overall | reasoning accuracy (%): 53.20 -> 61.67 | Improved |
| move from canonical edit prompt (Stage 1) to rephrased prompt (Stage 2) | attribute edits, averaged across models · UniKE attribute stages | text-side efficacy: reduces efficacy by roughly 70% | Hurt |
| move from Stage 1 to later stages | attribute edits, reasoning-augmented · UniKE attribute stages | VQA accuracy: decreases by roughly 35--40% | Hurt |
| compare PMET edit drift across models (conditioning pathway) | Ovis-U1 vs BLIP3o-4B · conditioning pathway analysis | relative drift at DiT input: 0.077 vs 0.458 (BLIP3o-4B) and 0.435 (OmniGen2) | Hurt |
Closest papers by meaning
- MMKE-Bench: A Multimodal Editing Benchmark for Diverse Visual Knowledge (ICLR 2025 · similarity 0.91)
- ComprehendEdit: A Comprehensive Dataset and Evaluation Framework for Multimodal Knowledge Editing (AAAI 2025 · similarity 0.91)
- ROVER: Benchmarking Reciprocal Cross-Modal Reasoning for Omnimodal Generation (ICLR 2026 · similarity 0.90)
- WEAVE: Unleashing and Benchmarking the In-context Interleaved Comprehension and Generation (CVPR 2026 · similarity 0.90)
- Towards Unified Multimodal Editing with Enhanced Knowledge Collaboration (NeurIPS 2024 · similarity 0.90)
- Factuality Matters: When Image Generation and Editing Meet Structured Visuals (ICLR 2026 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
