Research map / Label-efficient and robust learning / Knowledge distillation
The Bridge-Garden Dilemma in LLM Distillation: Why Mixing Hard and Soft Labels Works
Guanghui Wang, Kaiwen Lv Kacuila, Zhiyong Yang, Zitai Wang, Jin-Wen Wu, Longtao Huang, Qianqian Xu, Qingming Huang
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 |
|---|---|---|---|
| mix hard and soft labels with static weighting vs best pure KD | Qwen2.5-7B -> 3B distillation · general reasoning avg | accuracy avg: +1.69 | Improved |
| use hybrid (risk-guided) rather than best non-hybrid baseline | Forward KL distillation · Qwen2.5-3B general reasoning | accuracy avg: +1.95 | Improved |
| use soft KD instead of hard KD | Llama3.2-1B student · general reasoning avg | accuracy avg: 16.42 -> 23.25 | Improved |
| use hard KD instead of soft KD | Qwen2.5-3B student · general reasoning avg | accuracy avg: 51.87 -> 52.45 | Improved |
| hybrid hard-soft KD instead of on-policy KD | Qwen2.5-7B -> 3B · average accuracy | accuracy: +1.43% gain, 9.7x cheaper | Improved |
| add risk-guided hybrid over static weighting | Llama3.2-1B student · general reasoning avg | accuracy avg: 24.23 -> 26.27 | Improved |
| add curriculum scheduling over static weighting | Llama3.2-1B student · general reasoning avg | accuracy avg: 24.23 -> 25.67 | Improved |
| add confidence-based weighting over static weighting | Qwen2.5-3B student · ThmQA | accuracy: 25.45 -> 22.78 | Hurt |
| use intermediate mixing weight rather than pure hard or soft | Hybrid KD · multiple tasks | accuracy | Improved |
Closest papers by meaning
- Distilling What Matters: Confidence-Aware Selective Distillation for Large Language Models (NeurIPS 2026 · similarity 0.90)
- ABKD: Pursuing a Proper Allocation of the Probability Mass in Knowledge Distillation via $\alpha$-$\beta$-Divergence (ICML 2025 · similarity 0.90)
- Why Knowledge Distillation Works in Generative Models: A Minimal Working Explanation (NeurIPS 2025 · similarity 0.90)
- MiniPLM: Knowledge Distillation for Pre-training Language Models (ICLR 2025 · similarity 0.89)
- DistiLLM: Towards Streamlined Distillation for Large Language Models (ICML 2024 · similarity 0.89)
- DA-KD: Difficulty-Aware Knowledge Distillation for Efficient Large Language Models (ICML 2025 · similarity 0.89)
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.
