Research map / Vision-language and multimodal / Multimodal LLMs
Can Knowledge be Transferred from Unimodal to Multimodal? Investigating the Transitivity of Multimodal Knowledge Editing
Lingyong Fang, Xinzhong Wang, Depeng Wang, Zongru Wu, Ya Guo, Huijia Zhu, Zhuosheng Zhang, Gongshen Liu
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 |
|---|---|---|---|
| switch from unimodal to transitive editing (text edit, image test) | ROME · TMKEB LLaVA-1.5 | Reliability: 97.8 -> 30.6 | Hurt |
| switch from unimodal to transitive editing | SERAC · TMKEB LLaVA-1.5 | Reliability: 90.0 -> 26.1 | Hurt |
| switch from unimodal to transitive editing | MEND · TMKEB LLaVA-1.5 | Reliability: 98.6 -> 67.9 | Hurt |
| switch from unimodal to transitive editing | FT · TMKEB LLaVA-1.5 | Reliability: 100.0 -> 99.3 | No clear effect |
| switch from unimodal to transitive editing | IKE · TMKEB LLaVA-1.5 | Portability: 48.7 -> 24.9 | Hurt |
| edit with multimodal data instead of text (transitive to multimodal) | SERAC · TMKEB LLaVA-1.5 | Reliability: 26.1 -> 50.7 | Improved |
| edit with multimodal data instead of text | FT · TMKEB LLaVA-1.5 | Stability: 30.7 -> 6.5 | Hurt |
| edit with multimodal data instead of text | IKE · TMKEB LLaVA-1.5 | Stability: 20.2 -> 22.1 | No clear effect |
| use memory-based methods (SERAC, IKE) vs parameter-update methods | unimodal editing · TMKEB LLaVA-1.5 | Portability: 48.7 and 25.5 vs 20.3 | Improved |
| higher knowledge-neuron overlap M1 (D1 vs D2) (ablation) | ROME · TMKEB transitive | Reliability: 0.7574 vs 0.0152 | Improved |
| higher M2 overlap on Stability questions (D3 vs D4) (ablation) | ROME · TMKEB transitive | Stability: 0.1027 vs 0.6653 | Hurt |
| multimodal instead of transitive editing, M2 overlap (ablation) | FT · knowledge neurons | Stability: 0.1130 -> 0.0157 | Hurt |
Closest papers by meaning
- Towards Unified Multimodal Editing with Enhanced Knowledge Collaboration (NeurIPS 2024 · similarity 0.91)
- ComprehendEdit: A Comprehensive Dataset and Evaluation Framework for Multimodal Knowledge Editing (AAAI 2025 · similarity 0.91)
- MMKE-Bench: A Multimodal Editing Benchmark for Diverse Visual Knowledge (ICLR 2025 · similarity 0.89)
- MemEIC: A Step Toward Continual and Compositional Knowledge Editing (NeurIPS 2025 · similarity 0.87)
- Unveiling the Pitfalls of Knowledge Editing for Large Language Models (ICLR 2024 · similarity 0.86)
- Implicit Multimodal Alignment: On the Generalization of Frozen LLMs to Multimodal Inputs (NeurIPS 2024 · similarity 0.86)
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.
