Research map / Label-efficient and robust learning / Knowledge distillation
AMiD: Knowledge Distillation for LLMs with $\alpha$-mixture Assistant Distribution
Donghyeok Shin, Yeongmin Kim, Suhyeon Jo, Byeonghu Na, Il-chul Moon
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 |
|---|---|---|---|
| replace SRKL/TAID/ABKD with AMiD alpha-mixture assistant | GPT-2 XL -> GPT-2 0.1B · instruction-following avg | ROUGE-L: 21.76 -> 23.40 (vs ABKD) | Improved |
| AMiD vs ABKD on larger student | GPT-2 XL -> GPT-2 Large · instruction-following avg | ROUGE-L: 24.88 -> 25.84 | Improved |
| use small (negative) alpha with D_KL assistant (ablation) | AMiD, GPT-2 0.1B student · instruction-following avg | ROUGE-L: 18.16 (alpha=1.0) -> 22.66 (alpha=-5.0) | Improved |
| add assistant distribution to KL (ablation) | GPT-2 0.1B student · instruction-following avg | ROUGE-L: 17.61 -> 22.66 | Improved |
| use alpha=1.0 with D_RKL assistant (ablation) | AMiD, GPT-2 0.1B student · instruction-following avg | ROUGE-L: 21.46 -> 3.94 | Hurt |
| use alpha=1 (e-mixture) with D_AB (ablation) | AMiD, GPT-2 0.1B student · instruction-following avg | ROUGE-L: 21.76 -> 15.92 | Hurt |
| increase alpha (ablation) | AMiD · quality-diversity (ROUGE-L, Self-BLEU) | quality vs diversity | Mixed |
| raise lambda to 0.9 (ablation) | AMiD D_AB · instruction-following avg | ROUGE-L | Hurt |
| use lambda 0.5 instead of 0.1 (ablation) | AMiD D_AB · instruction-following avg | ROUGE-L | Improved |
| use alpha != +-1 instead of fixed alpha=-1 or 1 (ablation) | AMiD with fixed data · instruction-following avg | ROUGE-L: 19.32 (no assistant) -> 21.88 (Fixed data) | Improved |
| assistant with alpha=-1 on Gemma-2B translation and GSM8K (ablation) | AMiD, Gemma-7B -> 2B · Translation, GSM8K | COMET / accuracy: 74.21 -> 52.83 COMET; 24.26 -> 00.00 acc | Hurt |
| swap m-mixture alpha=-1 for alpha=-5 in DistiLLM-2 | DistiLLM-2, Qwen2.5-7B -> 1.5B · AlpacaEval, Evol-Inst, UltraFeed avg | win rate: 70.4 -> 71.7 | Improved |
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.
- ABKD — ABKD: Pursuing a Proper Allocation of the Probability Mass in Knowledge Distillation via $\alpha$-$\beta$-Divergence (ICML 2025) · better on 4 of 4 shared results
- TAID — TAID: Temporally Adaptive Interpolated Distillation for Efficient Knowledge Transfer in Language Models (ICLR 2025) · better on 3 of 3 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
- WASD: Wasserstein-based Knowledge Distillation for Large Language Models (NeurIPS 2026 · similarity 0.90)
- TAID: Temporally Adaptive Interpolated Distillation for Efficient Knowledge Transfer in Language Models (ICLR 2025 · similarity 0.90)
- 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)
- MiniPLM: Knowledge Distillation for Pre-training Language Models (ICLR 2025 · similarity 0.89)
- Multi-Level Optimal Transport for Universal Cross-Tokenizer Knowledge Distillation on Language Models (AAAI 2025 · similarity 0.88)
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.
