Research map / Label-efficient and robust learning / Knowledge distillation
Where Did This Sentence Come From? Tracing Provenance in LLM Reasoning Distillation
kaiyuan liu, Shaotian Yan, Rui Miao, Bing Wang, Chen Shen, Jun Zhang, Jieping Ye
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 |
|---|---|---|---|
| Select responses with most Teacher Sentences vs random selection | Qwen3-4B-Base distilled from DeepSeek-R1 · Average over AIME24, AIME25, MATH500, OlympiadBench | accuracy: 56.2 -> 58.7 | Improved |
| Select responses with most Teacher Sentences vs random selection | Qwen3-8B-Base · Average over AIME24, AIME25, MATH500, OlympiadBench | accuracy: 60.9 -> 62.6 | Improved |
| Select responses with most Teacher Sentences vs random selection | Qwen2.5-7B-Instruct (QwQ-32B teacher) · Average over AIME24, AIME25, MATH500, OlympiadBench | accuracy: 56.0 -> 57.7 | Improved |
| Select responses with most Teacher Sentences vs random selection | Qwen3-4B-Instruct-2507 (GPT-OSS-120B teacher) · Average over AIME24, AIME25, MATH500, OlympiadBench | accuracy: 74.0 -> 76.4 | Improved |
| Use GRAPE selection instead of random | Qwen3-4B-Base · Average over AIME24, AIME25, MATH500, OlympiadBench | accuracy: 56.2 -> 55.1 | Hurt |
| Select by relative proportion of Teacher Sentences instead of absolute count (ablation) | Qwen3-4B-Base · Average over AIME24, AIME25, MATH500, OlympiadBench | accuracy: 58.7 -> 56.4 | Hurt |
| Select responses with fewest Teacher Sentences (ablation) | Qwen3-4B-Base · Average over AIME24, AIME25, MATH500, OlympiadBench | accuracy: 58.7 -> 55.0 | Hurt |
| Select longest response instead of most Teacher Sentences (ablation) | Qwen3-4B-Base · Average over AIME24, AIME25, MATH500, OlympiadBench | accuracy: 58.7 -> 57.9 | Hurt |
| Use beta=0.15 or 0.05 instead of 0.1 (ablation) | Qwen3-4B-Base · Average over AIME24, AIME25, MATH500, OlympiadBench | accuracy: 58.7 -> 57.6 (0.15); 58.0 (0.05) | Hurt |
| Apply selection in science domain (OpenScienceReasoning_2) | Qwen3-4B-Base · GPQA-D | accuracy: 39.9 -> 42.9 | Improved |
| Apply selection in science domain | Qwen3-4B-Base · MATH500 | accuracy: 85.8 -> 87.4 | Improved |
Closest papers by meaning
- Distilling What Matters: Confidence-Aware Selective Distillation for Large Language Models (NeurIPS 2026 · similarity 0.89)
- Reasoning Scaffolding: Distilling the Flow of Thought from LLMs (ICLR 2026 · similarity 0.88)
- Collaborative Reasoning Distillation via Cross-Feedback and Coherent Curation (NeurIPS 2026 · similarity 0.88)
- DistiLLM-2: A Contrastive Approach Boosts the Distillation of LLMs (ICML 2025 · similarity 0.88)
- The Quest for Efficient Reasoning: A Data-Centric Benchmark to CoT Distillation (ICLR 2026 · similarity 0.88)
- Keypoint-based Progressive Chain-of-Thought Distillation for LLMs (ICML 2024 · 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.
