Research map / Large language models / Language, safety and interpretability
SelecTKD: Selective Token-Weighted Knowledge Distillation for LLMs
Haiduo Huang, Jiangcheng Song, Yadong Zhang, Pengju Ren
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 SelecTKD token weighting on top of DistiLLM-2 | DistiLLM-2, Qwen2-7B to 1.5B · Instruction following avg WR | win rate: 58.69 -> 59.47 | Improved |
| Add SelecTKD on top of vanilla KD | KD, Qwen2-7B to 1.5B · Instruction following avg WR | win rate: 41.19 -> 43.33 | Improved |
| Add SelecTKD on top of SKD | SKD, Mistral-7B to Danube2-1.8B · Evol-Instruct WR | win rate: 38.87 -> 38.96 | No clear effect |
| Add SelecTKD on top of DistiLLM-2 for math | DistiLLM-2, Qwen2-Math · GSM8K | Pass@1: 76.27 -> 77.82 | Improved |
| Add SelecTKD on top of DistiLLM-2 for code | DistiLLM-2, DS-Coder · HumanEval | Pass@1: 59.92 -> 61.72 | Improved |
| Distill a VLM with SelecTKD | InternVL2-2B · Open VLM Leaderboard average | average score: 54.0 -> 57.9 | Improved |
| Use Top-k/Spec-k verification instead of Hellinger-based weighting (ablation) | SelecTKD · Instruction following avg WR | win rate: 41.70 -> 43.33 | Improved |
| Use non-greedy Spec-k instead of greedy Top-k (ablation) | SelecTKD · Instruction following avg WR | win rate: 43.09 -> 43.33 | Improved |
| Give rejected tokens small weight beta=0.01 instead of masking (ablation) | SelecTKD Spec-k · Instruction following avg WR | win rate: 42.83 -> 43.33 | Improved |
| Increase k from 5 to 10 or 3/7 (ablation) | SelecTKD Spec-k · Instruction following avg WR | win rate: 43.33 -> 43.13 | No clear effect |
| Increase rejected-token weight beta to larger values (ablation) | SelecTKD · Instruction following avg WR | win rate: 43.33 -> 43.20 (beta 0.1) | Hurt |
| Vary loss function (KL/RKL/SKL/SRKL) inside SelecTKD (ablation) | SelecTKD · ROUGE-L on mixed data | ROUGE-L | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- DistiLLM-2 — DistiLLM-2: A Contrastive Approach Boosts the Distillation of LLMs (ICML 2025) · better on 9 of 9 shared results
- DistiLLM — DistiLLM: Towards Streamlined Distillation for Large Language Models (ICML 2024) · better on 1 of 1 shared result
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
- Distilling What Matters: Confidence-Aware Selective Distillation for Large Language Models (NeurIPS 2026 · similarity 0.93)
- WASD: Wasserstein-based Knowledge Distillation for Large Language Models (NeurIPS 2026 · similarity 0.90)
- DistiLLM: Towards Streamlined Distillation for Large Language Models (ICML 2024 · similarity 0.90)
- Hybrid Policy Distillation for LLMs (ICML 2026 · similarity 0.90)
- MiniPLM: Knowledge Distillation for Pre-training Language Models (ICLR 2025 · similarity 0.90)
- Entropy-Aware On-Policy Distillation of Language Models (ICML 2026 · 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.
