Research map / Label-efficient and robust learning / Knowledge distillation
Balanced Knowledge Distillation for Large Language Models with Mix-of-Experts
Jiajun Liu, Yao He, Wenjun Ke, Peng Wang, Ziyu Shang, Guozheng Li, Zijie Xu
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 Monte Carlo expert exploration (ablation) | B-Distill · General benchmarks avg (Llama3.2-1B-MoE) | Rouge-L: 22.80 -> 23.76 | Improved |
| add Monte Carlo expert exploration (ablation) | B-Distill · General benchmarks avg (Qwen3-0.6B-MoE) | Rouge-L: 12.96 -> 16.37 | Improved |
| add entropy-aware router distillation (ablation) | B-Distill · General benchmarks avg (Llama3.2-1B-MoE) | Rouge-L: 22.37 -> 23.76 | Improved |
| add entropy-aware router distillation (ablation) | B-Distill · General benchmarks avg (Qwen3-0.6B-MoE) | Rouge-L: 10.35 -> 16.37 | Improved |
| use B-Distill moe-to-moe instead of best dense-to-dense baseline | Llama3.2-1B-MoE · general avg Rouge-L | Rouge-L: 22.32 -> 23.76 | Improved |
| use B-Distill instead of best baseline | Qwen3-0.6B-MoE · medical/legal MMLU avg | accuracy: 0.42 -> 0.46 | Improved |
| increase exploration length K from 3 to above 3 (ablation) | B-Distill · general benchmarks Rouge-L | Rouge-L | Hurt |
| raise entropy weight beta from 0 to 0.1 (ablation) | B-Distill · general benchmarks avg | Rouge-L: 14.5 -> 16.0 (Llama), 21.5 -> 23.3 (Qwen) | Improved |
| raise entropy weight beta from 0.1 to 0.4 (ablation) | B-Distill · general benchmarks avg | Rouge-L: within 0.4 | No clear effect |
| use 4->4 experts instead of 4->1 (moe-to-dense) in student (ablation) | B-Distill · general benchmarks avg | Rouge-L: 20.9 -> 23.8 (Llama); 12.4 -> 15.6 (Qwen) | Improved |
Closest papers by meaning
- PADD: Path-Aligned Decompression Distillation for Non-Router Teacher to Guide MoE Student Learning (ICML 2026 · similarity 0.91)
- Layerwise Recurrent Router for Mixture-of-Experts (ICLR 2025 · similarity 0.90)
- Ada-K Routing: Boosting the Efficiency of MoE-based LLMs (ICLR 2025 · similarity 0.90)
- Advancing Expert Specialization for Better MoE (NeurIPS 2025 · similarity 0.89)
- $\textit{Read-ME}$: Refactorizing LLMs as Router-Decoupled Mixture of Experts with System Co-Design (NeurIPS 2024 · similarity 0.89)
- Distilling What Matters: Confidence-Aware Selective Distillation for Large Language Models (NeurIPS 2026 · 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.
