Research map / Vision-language and multimodal / Multimodal LLMs
Evaluating and Understanding Model Editing for Medical Vision Language Models
Guli Zhu, Chenwei Wu, Liyue Shen
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 |
|---|---|---|---|
| sequential instead of single editing | LoRA · M3Bench T-locality (LLaVA-Med 7B) | locality: 0.65 -> 0.03 | Hurt |
| sequential instead of single editing | BalancEdit · M3Bench Reliability (LLaVA-Med 7B) | reliability: 1.00 -> 0.60 | Hurt |
| use BalancEdit instead of LoRA for sequential editing | sequential editing · M3Bench T-locality (LLaVA-Med 7B) | locality: 0.03 -> 0.71 | Improved |
| use LoRA instead of BalancEdit for sequential editing | sequential editing · M3Bench C-generality (LLaVA-Med 7B) | compositional generality: 0.23 -> 0.68 | Improved |
| use BalancEdit instead of LoRA for sequential editing | sequential editing · M3Bench Overall (LLaVA-Med 7B) | harmonic mean: 0.21 -> 0.46 | Improved |
| use BalancEdit instead of LoRA for sequential editing | sequential editing · M3Bench Overall (Huatuo 34B) | harmonic mean: 0.35 -> 0.45 | Improved |
| sequential instead of single editing | LoRA · M3Bench Overall (LLaVA-Med 7B) | harmonic mean: 0.49 -> 0.21 | Hurt |
| add vision-language projector to edited scope (ablation) | BELoRA late LM layers · M3Bench harmonic mean of all tasks (Huatuo 7B) | harmonic mean: 0.45 -> 0.49 | Improved |
| add vision encoder to edited scope (ablation) | BELoRA late LM layers + projector · M3Bench harmonic mean of all tasks (Huatuo 7B) | harmonic mean: 0.49 -> 0.48 | No clear effect |
| edit full LM depth instead of late layers (ablation) | BELoRA LM · M3Bench harmonic mean of generality tasks (Huatuo 7B) | generality: 0.47 -> 0.36 | Hurt |
| edit full LM depth instead of late layers (ablation) | BELoRA LM · M3Bench harmonic mean of locality tasks (Huatuo 7B) | locality: 0.48 -> 0.52 | Improved |
| increase BalancEdit radius alpha from 0.01 to 0.20 (ablation) | BalancEdit · M3Bench Task1 Image Generality (LLaVA-Med) | performance: 0.37 -> 0.72 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- GRACE — GRACE: Graph-Based Contextual Debiasing for Fair Visual Question Answering (ECCV 2024) · better on 8 of 8 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- VLKEB: A Large Vision-Language Model Knowledge Editing Benchmark (NeurIPS 2024 · similarity 0.90)
- MMKE-Bench: A Multimodal Editing Benchmark for Diverse Visual Knowledge (ICLR 2025 · similarity 0.89)
- MultiMedBench: A Scenario-Aware Benchmark for Evaluating Knowledge Editing in Medical VQA (AAAI 2026 · similarity 0.88)
- Attribution Analysis Meets Model Editing: Advancing Knowledge Correction in Vision Language Models with VisEdit (AAAI 2025 · similarity 0.88)
- ImgEdit: A Unified Image Editing Dataset and Benchmark (NeurIPS 2025 · similarity 0.88)
- Uncovering and Mitigating Transient Blindness in Multimodal Model Editing (AAAI 2026 · similarity 0.88)
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.
