Research map / Label-efficient and robust learning / Knowledge distillation
GrayKD: Distilling Better Knowledge from Black-box LLM via Multi-rationale Injection
Hyeongsoo Lim, Hyung Yong Kim, Jin Young Kim, Min Ho Jang, Eun Seo Seo, Youshin Lim, Shukjae Choi, Jihwan Park, Yunkyu Lim, Hanbin Lee, Byeong-Yeol Kim, Ji Won Yoon
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 mix-sample strategy (ablation) | GrayKD (Triple) · 5 benchmarks avg | Rouge-L: 26.85 -> 27.64 | Improved |
| add rationale masking (ablation) | GrayKD (Triple) · 5 benchmarks avg | Rouge-L: 27.46 -> 27.64 | Improved |
| add both mix-sample and masking (ablation) | GrayKD (Triple) · 5 benchmarks avg | Rouge-L: 26.34 -> 27.64 | Improved |
| add masking (ablation) | GrayKD (Triple) · Dolly | Rouge-L: 26.66 -> 26.55 | No clear effect |
| use less diverse rationales (Sim 1, cosine 0.8618) (ablation) | GrayKD (Triple) · 5 benchmarks avg | Rouge-L: 27.64 -> 27.29 | Hurt |
| reuse a single rationale across all encoders (Sim 2) (ablation) | GrayKD (Triple) · 5 benchmarks avg | Rouge-L: 27.64 -> 26.50 | Hurt |
| use three rationales instead of two | GrayKD · 5 benchmarks avg | Rouge-L: 27.44 -> 27.64 | Improved |
| use GrayKD (Triple) instead of PromptKD with White Teacher 3 | SmolLM2-360M student · 5 benchmarks avg | Rouge-L: 26.44 -> 27.64 | Improved |
| distill with black-box SeqKD instead of white-box KD | SmolLM2-360M student · 5 benchmarks avg | Rouge-L: 26.11 -> 23.59 | Hurt |
| add ProxyKD to PromptKD with White Teacher 3 | SmolLM2-360M student · 5 benchmarks avg | Rouge-L: 26.44 -> 24.45 | Hurt |
| set lambda1=0.4 versus lambda1=0 (distillation only) (ablation) | GrayKD · dev set | Rouge-L | Improved |
| replace 1.7B white-box teacher training with rationale injection | GrayKD · total parameters used | parameters: 2.06B -> 610M | Improved |
Closest papers by meaning
- DA-KD: Difficulty-Aware Knowledge Distillation for Efficient Large Language Models (ICML 2025 · similarity 0.92)
- DistiLLM: Towards Streamlined Distillation for Large Language Models (ICML 2024 · similarity 0.91)
- DDK: Distilling Domain Knowledge for Efficient Large Language Models (NeurIPS 2024 · similarity 0.91)
- TAID: Temporally Adaptive Interpolated Distillation for Efficient Knowledge Transfer in Language Models (ICLR 2025 · similarity 0.90)
- WASD: Wasserstein-based Knowledge Distillation for Large Language Models (NeurIPS 2026 · 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.
