Research map / Vision-language and multimodal / Multimodal LLMs
VisNec: Measuring and Leveraging Visual Necessity for Multimodal Instruction Tuning
Mingkang Dong, Hongyi Cai, jie li, Sifan Zhou, Bin Ren, Kunyu Peng, Yuqian Fu
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 |
|---|---|---|---|
| select data with VisNec (blind vs multimodal loss difference plus clustering) instead of random | LLaVA-v1.5-7B 15% subset · LLaVA-665K | relative performance: 94.2% -> 100.2% | Improved |
| select data with VisNec instead of random | LLaVA-v1.5-7B 15% subset · Vision-Flan-186K | relative performance: 96.7% -> 115.8% | Improved |
| apply VisNec selection instead of random on Qwen2.5-VL-3B | Qwen2.5-VL-3B 98K subset · LLaVA-665K | relative performance: 98.8% -> 103.8% | Improved |
| apply VisNec selection instead of random on Qwen2.5-VL-7B | Qwen2.5-VL-7B 98K subset · LLaVA-665K | relative performance: 98.7% -> 104.0% | Improved |
| apply VisNec selection instead of random on Qwen2.5-VL-32B | Qwen2.5-VL-32B 98K subset · LLaVA-665K | relative performance: 97.9% -> 102.4% | Improved |
| use the loss difference (blind minus multimodal) instead of text-only loss alone (ablation) | VisNec · LLaVA-665K 15% | relative performance: 95.6% -> 100.2% | Improved |
| use the loss difference instead of multimodal loss alone (ablation) | VisNec · LLaVA-665K 15% | relative performance: 94.0% -> 100.2% | Improved |
| add semantic clustering (K=20) before intra-cluster top-r% selection (ablation) | VisNec · LLaVA-665K 15% | relative performance: 97.0% -> 100.2% | Improved |
| vary number of clusters k from 10 to 30 (ablation) | VisNec · LLaVA-665K 15% | relative performance: fluctuates within 1.3%, over 98.9% | No clear effect |
| increase selection ratio from 15% to 20% (ablation) | VisNec · LLaVA-665K | relative performance: 100.2% -> 99.2% | Hurt |
| reduce selection ratio to 5% (ablation) | VisNec · LLaVA-665K | relative performance: 100.2% -> 99.1% | Mixed |
| use VisNec instead of Self-Filter selection for cost | VisNec · LLaVA-665K | selection time: 73.5 GPU-hr -> 12.0 GPU-hr | 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.
- CoIDO — CoIDO: Efficient Data Selection for Visual Instruction Tuning via Coupled Importance-Diversity Optimization (NeurIPS 2025) · better on 2 of 2 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
- Adapt-$\infty$: Scalable Continual Multimodal Instruction Tuning via Dynamic Data Selection (ICLR 2025 · similarity 0.90)
- Learning to Instruct for Visual Instruction Tuning (NeurIPS 2025 · similarity 0.90)
- INST-IT: Boosting Instance Understanding via Explicit Visual Prompt Instruction Tuning (NeurIPS 2025 · similarity 0.88)
- Improved Baselines with Visual Instruction Tuning (CVPR 2024 · similarity 0.87)
- Re-Imagining Multimodal Instruction Tuning: A Representation View (ICLR 2025 · similarity 0.87)
- VIGC: Visual Instruction Generation and Correction (AAAI 2024 · similarity 0.87)
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.
