Research map / Vision-language and multimodal / Multimodal LLMs
LLaVA-KD: A Framework of Distilling Multimodal Large Language Models
Yuxuan Cai, Jiangning Zhang, Haoyang He, Xinwei He, Ao Tong, Zhenye Gan, Chengjie Wang, Zhucun Xue, Yong Liu, Xiang Bai
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 distilled pre-training (DPT) before SFT (ablation) | PT-SFT Qwen1.5-0.5B · Avg10 (10 benchmarks) | Avg10: 54.7 -> 55.6 | Improved |
| add distilled fine-tuning (DFT) after DPT-SFT (ablation) | DPT-SFT · Avg10 (10 benchmarks) | Avg10: 55.6 -> 57.9 | Improved |
| add SFT stage between DPT and DFT (ablation) | DPT-DFT · Avg10 (10 benchmarks) | Avg10: 55.9 -> 57.9 | Improved |
| use DPT instead of plain PT in the SFT-DFT pipeline (ablation) | PT-SFT-DFT · Avg10 (10 benchmarks) | Avg10: 56.6 -> 57.9 | Improved |
| use DPT instead of plain PT in the no-SFT pipeline (ablation) | PT-DFT · Avg10 (10 benchmarks) | Avg10: 55.8 -> 55.9 | No clear effect |
| add a second DFT stage (DPT-DFT-DFT) (ablation) | DPT-SFT-DFT · Avg10 (10 benchmarks) | Avg10: 57.9 -> 58.0 | No clear effect |
| add RDist to MDist during DPT (ablation) | DPT MDist only · Avg10 (10 benchmarks) | Avg10: 55.1 -> 55.6 | Improved |
| add RDist to MDist during DFT (ablation) | DFT MDist only · Avg10 (10 benchmarks) | Avg10: 57.7 -> 57.9 | Improved |
| add MDist to RDist during DFT (ablation) | DFT RDist only · Avg10 (10 benchmarks) | Avg10: 57.0 -> 57.9 | Improved |
| add visual-token distillation to response distillation in DFT (ablation) | MDist DFT · Avg10 (10 benchmarks) | Avg10: 57.2 -> 57.7 | Improved |
| also distill prompt tokens together with response and visual tokens in DFT (ablation) | MDist DFT · Avg10 (10 benchmarks) | Avg10: 57.7 -> 57.1 | Hurt |
| use MDist+RDist instead of FKL/JSD under PT-DFT (ablation) | PT-DFT · Avg10 (10 benchmarks) | Avg10: 55.5 -> 55.8 | Improved |
Later papers that compared with this method
1 later accepted paper reports LLaVA-KD as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Beyond Next-Token Alignment: Distilling Multimodal Large Language Models via Token Interactions (ICML 2026 · 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
- Predictive Regularization Against Visual Representation Degradation in Multimodal Large Language Models (CVPR 2026 · similarity 0.91)
- NaViL: Rethinking Scaling Properties of Native Multimodal Large Language Models under Data Constraints (NeurIPS 2025 · similarity 0.91)
- Unleashing the Intrinsic Visual Representation Capability of Multimodal Large Language Models (CVPR 2026 · similarity 0.91)
- LLaDA-V: Large Language Diffusion Models with Visual Instruction Tuning (CVPR 2026 · similarity 0.91)
- LRM-LLaVA: Overcoming the Modality Gap of Multilingual Large Language-Vision Model for Low-Resource Languages (AAAI 2025 · similarity 0.91)
- Beyond Next-Token Alignment: Distilling Multimodal Large Language Models via Token Interactions (ICML 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.
