Research map / Label-efficient and robust learning / Knowledge distillation
Distillation of Large Language Models via Concrete Score Matching
Yeongmin Kim, Donghyeok Shin, Mina Kang, Byeonghu Na, Il-chul Moon
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 CSD instead of KL for distillation | GPT-2-1.5B -> GPT-2-0.1B · instruction following | avg ROUGE-L: 16.68 -> 20.65 | Improved |
| use CSD instead of DLD with student weighting (ablation) | GPT-2-0.1B student · instruction following | avg ROUGE-L: 20.39 -> 20.65 | Improved |
| use CSD instead of DLD with uniform weighting (ablation) | GPT-2-0.1B student · instruction following | avg ROUGE-L: 8.63 (DLD U) vs 13.37 (CSD U,U) | Mixed |
| add CSD loss to ImitKD (on-policy) | ImitKD, GPT-2-0.1B · instruction following | avg ROUGE-L: 16.69 -> 20.73 | Improved |
| replace DLD with CSD inside ImitKD | ImitKD, GPT-2-0.3B · instruction following | avg ROUGE-L: 22.32 -> 23.12 | Improved |
| add CSD to DistiLLM | DistiLLM, GPT-2-0.3B · instruction following | avg ROUGE-L: 22.67 -> 22.72 | No clear effect |
| scale CSD with ImitKD to OpenLLaMA-7B -> 3B | OpenLLaMA-3B · instruction following | avg ROUGE-L: 28.31 -> 29.03 | Improved |
| use CSD (T,S) instead of KL on GSM8K distillation | Gemma-2B-IT · GSM8K | accuracy: 24.03 -> 25.78 | Improved |
| use CSD (S,S) student weighting in low-data task distillation (ablation) | Gemma-2B-IT · GSM8K | accuracy: 25.78 -> 0.00 | Hurt |
| use uniform or teacher weight for w1 instead of student (ablation) | CSD, GPT-2-0.1B · instruction following | avg ROUGE-L: 20.65 -> 19.88 (U,S), 17.59 (T,S) | Hurt |
| use teacher weights for both w1 and w2 (ablation) | CSD, GPT-2-0.1B · instruction following | avg ROUGE-L: 20.65 -> 5.91 | Hurt |
| use CSD (S,S) instead of DistiLLM-2 for chat distillation | Qwen2.5-1.5B-IT · MT-Bench (GPT4 judge) | MT-Bench score: 7.28 -> 7.69 | Improved |
Closest papers by meaning
- Distilling What Matters: Confidence-Aware Selective Distillation for Large Language Models (NeurIPS 2026 · similarity 0.91)
- DistiLLM: Towards Streamlined Distillation for Large Language Models (ICML 2024 · similarity 0.91)
- WASD: Wasserstein-based Knowledge Distillation for Large Language Models (NeurIPS 2026 · similarity 0.91)
- DA-KD: Difficulty-Aware Knowledge Distillation for Efficient Large Language Models (ICML 2025 · similarity 0.90)
- MiniPLM: Knowledge Distillation for Pre-training Language Models (ICLR 2025 · similarity 0.90)
- DDK: Distilling Domain Knowledge for Efficient Large Language Models (NeurIPS 2024 · 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.
