Research map / Label-efficient and robust learning / Knowledge distillation
Don't Ignore the Tail: Decoupled Distillation Produces Top Maths Students on an Academic Budget
Sayantan Dasgupta, Trevor Cohn, Timothy Baldwin
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 KL with tail-aware divergence (K=10) | pretraining distillation, Qwen1.5-1.8B -> 1.2B · 8-task avg Regmix | average accuracy (%): 45.6 -> 47.8 | Improved |
| replace KL with tail-aware divergence (K=10) | Phi2 2.8B -> 1.1B · 8-task avg Regmix | average accuracy (%): 49.1 -> 50.3 | Improved |
| replace KL with tail-aware divergence (K=10) | Gemma2-9B -> 2B · 8-task avg Regmix | average accuracy (%): 45.8 -> 47.0 | Improved |
| replace KL with tail-aware divergence (K=10) | Qwen2.5-3B -> 1.2B · 8-task avg Regmix | average accuracy (%): 44.9 -> 46.3 | Improved |
| use MiniPLM data selection instead of vanilla KD | Phi2 2.8B -> 1.1B · 8-task avg Regmix | average accuracy (%): 49.1 -> 48.8 | Hurt |
| increase top-K beyond 10 (ablation) | TAD · 8-task avg Regmix | average accuracy (%): 47.8 -> 47.7 | No clear effect |
| vary beta (0.5 to 10) with K=10 (ablation) | TAD, Qwen1.8B -> 1.2B · 8-task avg Regmix | average accuracy (%): 47.0 / 47.6 / 47.8 / 47.7 / 47.6 | No clear effect |
| add cosine loss on hidden states | Vanilla KD and TAD, larger teachers · 8-task avg Regmix | average accuracy (%) | Improved |
| use reverse KL instead of forward KL (ablation) | distillation with cosine loss · Full-ECE Regmix | F-ECE: 1.45 -> 1.77 | Hurt |
| use TAD instead of vanilla KD | Phi2 2.8B -> 1.1B · Full-ECE Regmix | F-ECE: 1.45 -> 1.19 | Improved |
| increase K from 1 to 20 (ablation) | TAD · Full-ECE | F-ECE: 1.19 -> 1.42 | Hurt |
| use MiniPLM for domain-specific distillation | TinyLlama-1.1B on OWM · math 8-task avg | average accuracy (%): 16.7 -> 14.4 | 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.
- MiniPLM — MiniPLM: Knowledge Distillation for Pre-training Language Models (ICLR 2025) · better on 6 of 6 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
- Distilling What Matters: Confidence-Aware Selective Distillation for Large Language Models (NeurIPS 2026 · similarity 0.91)
- ABKD: Pursuing a Proper Allocation of the Probability Mass in Knowledge Distillation via $\alpha$-$\beta$-Divergence (ICML 2025 · similarity 0.91)
- Enhancing Logits Distillation with Plug&Play Kendall's $\tau$ Ranking Loss (ICML 2025 · similarity 0.90)
- Entropy-Aware On-Policy Distillation of Language Models (ICML 2026 · similarity 0.89)
- DistiLLM: Towards Streamlined Distillation for Large Language Models (ICML 2024 · similarity 0.89)
- PLD: A Choice-Theoretic List-Wise Knowledge Distillation (NeurIPS 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.
