Research map / Label-efficient and robust learning / Knowledge distillation
Boomerang Distillation Enables Zero-Shot Model Size Interpolation
Sara Kangaslahti, Nihal V. Nayak, Jonathan Geuter, Marco Fumero, Francesco Locatello, David Alvarez-Melis
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 |
|---|---|---|---|
| initialize student from teacher weights instead of randomly (ablation) | Boomerang distillation, Qwen3-4B-Base · 10 classification and 3 generation tasks | accuracy vs parameter count | Improved |
| add knowledge distillation training to the layer-pruned student (versus naive layer pruning) (ablation) | Boomerang distillation, Qwen3-4B-Base · classification and generation tasks | accuracy | Improved |
| add per-layer cosine alignment loss to CE+KL (ablation) | Boomerang distillation, Qwen3-4B-Base · WikiText | perplexity of interpolated models | Improved |
| add per-layer alignment loss (ablation) | Boomerang distillation, Qwen3-4B-Base · downstream classification/generation, majority of sizes | accuracy | No clear effect |
| add per-layer alignment loss (ablation) | Boomerang distillation, Qwen3-4B-Base · extreme intermediate sizes (first/last layers patched) | stability of accuracy | Improved |
| patch teacher layers into the student (boomerang) versus layer pruning LaCo and ShortGPT | Qwen3-4B-Base interpolated models · classification and generation tasks | accuracy | Improved |
| compare interpolated models to standard distilled models of same size | Boomerang distillation, Qwen3-4B-Base · classification and generation tasks | accuracy | Mixed |
| compare interpolated models to standard distilled models on Pythia-2.8B | Boomerang distillation, Pythia-2.8B · classification tasks | accuracy | Hurt |
| use student with every 3rd layer kept instead of every 2nd (ablation) | Boomerang distillation, Qwen3-4B-Base · student size ablation | interpolation quality | No clear effect |
| keep every 4th or 5th layer for student init instead of every 2nd (ablation) | Boomerang distillation, Qwen3-4B-Base · student size ablation | interpolation smoothness | Hurt |
| choose student layers to preserve first/last cosine similarity in each teacher block (ablation) | Boomerang distillation, Qwen3-4B-Base · small student models | interpolation smoothness: 505M parameters, 8.7x compression | Improved |
| increase student training tokens from 0.5B up to 3.1B (ablation) | Boomerang distillation, Qwen3-4B-Base · token budget ablation | interpolated model performance: 0.5B, 1B, 1.5B, 2B, 2.5B, 3.1B tokens | Improved |
Closest papers by meaning
- Understanding Layer Patching in Model Size Interpolation (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)
- Distilled Pretraining: A modern lens of Data, In-Context Learning and Test-Time Scaling (ICLR 2026 · similarity 0.88)
- DistiLLM-2: A Contrastive Approach Boosts the Distillation of LLMs (ICML 2025 · similarity 0.88)
- Model Merging in Pre-training of Large Language Models (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.
