Research map / Large language models / Language, safety and interpretability
A Unified Approach to Interpreting Knowledge Distillation for Large Language Models via Interactions
Qingzhuo Wang, Ruiyang Qin, Zhenxin Qin, Wen Shen, Zhihua Wei
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 |
|---|---|---|---|
| distill the student with KD instead of leaving it undistilled (interaction sparsity) | GPT-2, OPT, LLaMA students · Dolly test interactions | Gini coefficient / entropy | Improved |
| standard SFT instead of KD (ablation) | student models · Dolly test interactions | interaction sparsity | No clear effect |
| apply KD to the student | GPT-2-0.1B student · Dolly test, overlap@0.1 with teacher | overlap rate: 19.81% -> 24.84% (KD) | Improved |
| apply DISTILLM to LLaMA-7B student | LLaMA-7B · Dolly test, overlap@0.1 | overlap rate: 21.35% -> 31.60% | Improved |
| higher sparsity of complex interactions across KD methods | KD methods · Dolly, GPT-2 and OPT students | ROUGE-L correlation: Pearson 0.81 (+-0.11) complex vs 0.34 (+-0.12) simple | Improved |
| add Complex Interaction Penalty (CIP) loss to GKD | GKD, GPT-2-0.3B · Dolly test | ROUGE-L: 26.35 -> 26.52 | Improved |
| add CIP loss to DISTILLM | DISTILLM, GPT-2-0.3B · Dolly test | ROUGE-L: 27.47 -> 27.67 | Improved |
| add CIP loss to GKD | GKD, OPT-1.3B · Dolly test | ROUGE-L: 26.00 -> 26.38 | Improved |
| add CIP loss to DISTILLM | DISTILLM, OPT-1.3B · Dolly test | ROUGE-L: 27.69 -> 27.77 | Improved |
| add generic L1 penalty instead of CIP (ablation) | GKD and DISTILLM · Dolly test, GPT-2-0.3B and OPT-1.3B | ROUGE-L: -1.17 to -0.56 | Hurt |
| add Tf-KD regularization instead of CIP (ablation) | GKD and DISTILLM · Dolly test, GPT-2-0.3B and OPT-1.3B | ROUGE-L: -0.47 to -0.04 | Hurt |
| increase CIP weight lambda (ablation) | CIP on GPT-2-0.3B · Dolly test | ROUGE-L | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DISTILLM — DistiLLM: Towards Streamlined Distillation for Large Language Models (ICML 2024) · better on 2 of 2 shared results
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
- Why Knowledge Distillation Works in Generative Models: A Minimal Working Explanation (NeurIPS 2025 · similarity 0.91)
- DistiLLM: Towards Streamlined Distillation for Large Language Models (ICML 2024 · similarity 0.90)
- DA-KD: Difficulty-Aware Knowledge Distillation for Efficient Large Language Models (ICML 2025 · similarity 0.89)
- DDK: Distilling Domain Knowledge for Efficient Large Language Models (NeurIPS 2024 · similarity 0.89)
- WASD: Wasserstein-based Knowledge Distillation for Large Language Models (NeurIPS 2026 · similarity 0.89)
- Multi-Level Optimal Transport for Universal Cross-Tokenizer Knowledge Distillation on Language Models (AAAI 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.
