Research map / Large language models / Instruction and fine-tuning
How to Fine-Tune a Reasoning Model? A Teacher–Student Cooperation Framework to Synthesize Student-Consistent SFT Data
Zixian Huang, Kaichen Yang, Xu Huang, Feiyang Hao, Qiming Ge, Bowen Li, He Du, Kai Chen, Qipeng Guo
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 |
|---|---|---|---|
| use TESSY teacher-student cooperative synthesis instead of Teacher-Only data | Qwen3-8B SFT · OJBench | pass@1: 8.73 -> 25.43 | Improved |
| use TESSY instead of Teacher-Only data | Qwen3-8B SFT · LiveCodeBench-Pro | pass@1: 22.10 -> 36.69 | Improved |
| SFT on Teacher-Only (GPT-OSS-120B) data | Qwen3-8B · OJBench | pass@1: 18.75 -> 8.73 | Hurt |
| SFT on Self-Distillation data | Qwen3-8B · LiveCodeBench-Pro | pass@1: 25.35 -> 19.26 | Hurt |
| SFT on Reject-Sampling data | Qwen3-8B · LiveCodeBench-Pro | pass@1: 25.35 -> 26.06 | No clear effect |
| use TESSY instead of Teacher-Only with DeepSeek-R1-Distill-Llama-8B student | DeepSeek-R1-Distill-Llama-8B · LiveCodeBench-V5 | pass@1: 38.32 -> 52.10 | Improved |
| use student (not teacher) to generate final answer in TESSY (ablation) | TESSY · LiveCodeBench-Pro / OJBench | pass@1: 12.33% and 13.58% drops when teacher generates answer | Improved |
| train boundary predictors on 100K instead of 500 samples (ablation) | TESSY · OJBench (40K samples) | pass@1: 23.98 -> 26.51 | Improved |
| add independent system prompt isolation to Teacher-Only training (ablation) | Teacher-Only SFT · OJBench / LiveCodeBench-Pro | pass@1: +2.55% and +5.28% | Improved |
| use LoRA (rank 8, alpha 16) instead of full fine-tuning (ablation) | Qwen3-8B SFT · LiveCodeBench-Pro / OJBench | pass@1 | Hurt |
| start SFT from reasoning model Qwen3-8B instead of Qwen3-8B-Base (ablation) | TESSY SFT · OJBench / LiveCodeBench-Pro | pass@1: base lags by 17.67% | Improved |
| use TESSY on Qwen3-30B-A3B student | Qwen3-30B-A3B · LiveCodeBench-Pro / OJBench | pass@1: +6.37% and +10.13% | Improved |
Closest papers by meaning
- The Quest for Efficient Reasoning: A Data-Centric Benchmark to CoT Distillation (ICLR 2026 · similarity 0.88)
- Collaborative Reasoning Distillation via Cross-Feedback and Coherent Curation (NeurIPS 2026 · similarity 0.87)
- Advancing Mathematical Reasoning in Language Models: The Impact of Problem-Solving Data, Data Synthesis Methods, and Training Stages (ICLR 2025 · similarity 0.86)
- The Signal is in the Steps: Local Scoring for Reasoning Data Selection (ICML 2026 · similarity 0.86)
- RLKD: Distilling LLMs’ Reasoning via Reinforcement Learning (AAAI 2026 · similarity 0.86)
- UFT: Unifying Supervised and Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.86)
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.
