Research map / Label-efficient and robust learning / Knowledge distillation
EM-KD: Distilling Efficient Multimodal Large Language Model with Unbalanced Vision Tokens
Ze Feng, Sen Yang, Boqiang Duan, Wankou Yang, Jingdong Wang
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 vision-language affinity distillation (VLAD) to the SFT baseline (ablation) | 0.6B efficient student · 11-benchmark average | average accuracy: 47.7 -> 48.4 | Improved |
| add vision semantic distillation (VSD) on top of VLAD (ablation) | 0.6B efficient student · 11-benchmark average | average accuracy: 48.4 -> 49.5 | Improved |
| add response logits distillation (RLD) on top of VLAD and VSD (ablation) | 0.6B efficient student · 11-benchmark average | average accuracy: 49.5 -> 50.4 | Improved |
| add VLAD alone (ablation) | 0.6B efficient student · Knowledge benchmarks average | accuracy: 42.4 -> 41.1 | Hurt |
| match tokens with Hungarian algorithm on vision logits instead of average pooling (ablation) | EM-KD · 11-benchmark average | average accuracy: 48.6 -> 50.4 | Improved |
| match with vision logits instead of hidden states in Hungarian matching (ablation) | EM-KD · 11-benchmark average | average accuracy: 49.4 -> 50.4 | Improved |
| distill vision logits instead of hidden states in VSD (ablation) | EM-KD · 11-benchmark average | average accuracy: 49.5 -> 50.4 | Improved |
| use EM-KD instead of LLaVA-KD with Vision Token Matching | 0.6B efficient student · 11-benchmark average | average accuracy: 49.7 -> 50.4 | Improved |
| use EM-KD instead of MiniLLM response-only distillation | 0.6B efficient student · 11-benchmark average | average accuracy: 49.5 -> 50.4 | Improved |
| distill from LLaVA-OneVision-SI-7B teacher into 8B efficient student with EM-KD | 8B efficient student · 11-benchmark average | average accuracy: 61.7 -> 63.4 | Improved |
| train student with EM-KD instead of training-free pruning or compression at 144 tokens | LLaVA-NeXT 0.5B · RealWorldQA TTFT | TTFT (ms): 103.3 -> 54.9 | Improved |
Later papers that compared with this method
1 later accepted paper reports EMKD as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Bidirectional Counterfactual Distillation for Review-Based Recommendation (AAAI 2026 · better on 5 of 5 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
- Beyond Next-Token Alignment: Distilling Multimodal Large Language Models via Token Interactions (ICML 2026 · similarity 0.90)
- Align-KD: Distilling Cross-Modal Alignment Knowledge for Mobile Vision-Language Large Model Enhancement (CVPR 2025 · similarity 0.90)
- MoVE-KD: Knowledge Distillation for VLMs with Mixture of Visual Encoders (CVPR 2025 · similarity 0.88)
- Hawaii: Hierarchical Visual Knowledge Transfer for Efficient Vision-Language Models (NeurIPS 2025 · similarity 0.88)
- LION: Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge (CVPR 2024 · similarity 0.88)
- No Head Left Behind – Multi-Head Alignment Distillation for Transformers (AAAI 2024 · 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.
