Research map / Label-efficient and robust learning / Knowledge distillation
Knowledge Distillation for Large Language Models through Residual Learning
Thinh On, Hengzhi Pei, Leonard Lausen, George Karypis
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 scaling factor beta in residual computation (ablation) | residual learning · Instruction following avg, Mixtral-8x7B -> GPT2-120M (cross-tokenizer) | Avg Rouge-L: 16.25 -> 20.01 | Improved |
| add teacher-error accuracy mask (ablation) | residual learning · Instruction following avg, Mixtral-8x7B -> GPT2-120M (cross-tokenizer) | Avg Rouge-L: 19.59 -> 20.01 | Improved |
| add pretraining of teacher projector (ablation) | residual learning · Instruction following avg, Mixtral-8x7B -> GPT2-120M (cross-tokenizer) | Avg Rouge-L: 17.39 -> 20.01 | Improved |
| freeze the pretrained teacher projector during distillation (ablation) | residual learning · Instruction following avg, Mixtral-8x7B -> GPT2-120M (cross-tokenizer) | Avg Rouge-L: 19.07 -> 20.01 | Improved |
| add expert knowledge fusion instead of sparse 2/8 activation (ablation) | MoE teacher · Instruction following avg, Mixtral-8x7B -> GPT2-120M (cross-tokenizer) | Avg Rouge-L: 19.66 -> 20.01 | Improved |
| add expert knowledge fusion instead of dense 8/8 activation (ablation) | MoE teacher · Instruction following avg, Mixtral-8x7B -> GPT2-120M (cross-tokenizer) | Avg Rouge-L: 19.68 -> 20.01 | Improved |
| add expert knowledge fusion instead of average pooling (ablation) | MoE teacher · Instruction following avg, Mixtral-8x7B -> GPT2-120M (cross-tokenizer) | Avg Rouge-L: 18.43 -> 20.01 | Improved |
| increase alignment dimension dA from 64 to 768 (ablation) | residual learning · Instruction following avg, Mixtral-8x7B -> GPT2-120M (cross-tokenizer) | Avg Rouge-L: 20.01 -> 18.02 | Hurt |
| add residual learning to DSKD (ablation) | DSKD · Instruction following avg, Mixtral-8x7B -> GPT2-120M (cross-tokenizer) | Avg Rouge-L: 18.40 -> 19.60 | Improved |
| use residual learning instead of DSKD with reverse KL (ablation) | DSKD · Instruction following avg, Mixtral-8x7B -> GPT2-120M (cross-tokenizer) | Avg Rouge-L: 15.98 -> 20.01 | Improved |
| lower lambda from 0.5 to 0.1 (ablation) | residual learning · Instruction following avg, Mixtral-8x7B -> GPT2-120M (cross-tokenizer) | Avg Rouge-L: 20.01 -> 16.39 | Hurt |
| raise lambda from 0.5 to 0.9 (ablation) | residual learning · Instruction following avg, Mixtral-8x7B -> GPT2-120M (cross-tokenizer) | Avg Rouge-L: 20.01 -> 18.81 | Hurt |
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.
- ABKD — ABKD: Pursuing a Proper Allocation of the Probability Mass in Knowledge Distillation via $\alpha$-$\beta$-Divergence (ICML 2025) · 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
- DistiLLM: Towards Streamlined Distillation for Large Language Models (ICML 2024 · similarity 0.92)
- TAID: Temporally Adaptive Interpolated Distillation for Efficient Knowledge Transfer in Language Models (ICLR 2025 · similarity 0.91)
- WASD: Wasserstein-based Knowledge Distillation for Large Language Models (NeurIPS 2026 · similarity 0.91)
- DDK: Distilling Domain Knowledge for Efficient Large Language Models (NeurIPS 2024 · similarity 0.91)
- Multi-Level Optimal Transport for Universal Cross-Tokenizer Knowledge Distillation on Language Models (AAAI 2025 · similarity 0.90)
- MiniPLM: Knowledge Distillation for Pre-training Language Models (ICLR 2025 · 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.
