Research map / Label-efficient and robust learning / Knowledge distillation
MCW-KD: Multi-Cost Wasserstein Knowledge Distillation for Large Language Models
Hoang Tran Vuong, Tue Le, Quyen Tran, Linh Ngo Van, Trung Le
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 |
|---|---|---|---|
| use MCW-KD instead of DSKD | student LLM distillation · Qwen2.5-7B-Instruct -> GPT2-1.5B average | Rouge-L (%): 19.27 -> 21.66 | Improved |
| use MCW-KD instead of DSKD | student LLM distillation · Mistral-7B -> TinyLLaMA-1.1B average | Rouge-L (%): 21.33 -> 22.79 | Improved |
| add multi-cost hidden-state alignment to DSKD (ablation) | DSKD · Mistral-7B -> TinyLLaMA-1.1B average | Rouge-L (%): 21.33 -> 21.65 | Improved |
| add multi-cost output alignment to DSKD (ablation) | DSKD · Qwen2.5-7B-Instruct -> GPT2-1.5B average | Rouge-L (%): 19.27 -> 20.42 | Improved |
| add both alignment spaces with learned dynamic weights instead of fixed 1/3 weights (ablation) | MCW-KD · Qwen2.5-7B-Instruct -> GPT2-1.5B average | Rouge-L (%): 20.90 -> 21.66 | Improved |
| add both alignment spaces with learned dynamic weights instead of fixed 1/3 weights (ablation) | MCW-KD · Mistral-7B -> TinyLLaMA-1.1B average | Rouge-L (%): 21.60 -> 22.79 | Improved |
| add MCW-KD loss to existing KD baselines | ULD, MultiLevelOT, MinED · Qwen1.5-1.8B -> GPT2-120M average | Rouge-L (%): ULD 14.01 -> 15.16; MultiLevelOT 13.25 -> 14.01; MinED 15.08 -> 15.72 | Improved |
| add MCW-KD loss to MultiLevelOT | MultiLevelOT · Dolly (Qwen1.5-1.8B -> GPT2-120M) | Rouge-L (%): 23.02 -> 22.46 | Hurt |
| vary gamma over 0.2-10 for the DSKD loss weight (ablation) | MCW-KD · Qwen2.5-7B-Instruct -> GPT2-1.5B | Rouge-L (%) | No clear effect |
| vary lambda over 0.2-5 (ablation) | MCW-KD · Qwen1.5-1.8B -> GPT2-340M | Rouge-L (%) | No clear effect |
| use MCW-KD instead of DSKD with shared vocabulary | student LLM distillation · GPT2-1.5B -> GPT2-120M average | Rouge-L (%): 16.60 -> 17.31 | Improved |
Closest papers by meaning
- WASD: Wasserstein-based Knowledge Distillation for Large Language Models (NeurIPS 2026 · similarity 0.95)
- Multi-Level Optimal Transport for Universal Cross-Tokenizer Knowledge Distillation on Language Models (AAAI 2025 · similarity 0.92)
- DistiLLM: Towards Streamlined Distillation for Large Language Models (ICML 2024 · similarity 0.92)
- DA-KD: Difficulty-Aware Knowledge Distillation for Efficient Large Language Models (ICML 2025 · similarity 0.91)
- DDK: Distilling Domain Knowledge for Efficient Large Language Models (NeurIPS 2024 · similarity 0.90)
- TAID: Temporally Adaptive Interpolated Distillation for Efficient Knowledge Transfer in Language Models (ICLR 2025 · 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.
