How to Fine-Tune a Reasoning Model? A Teacher-Student Cooperation Framework to Synthesize Student-Consistent SFT Data
For practitioners fine-tuning reasoning models, TESSY offers a method to effectively leverage stronger teachers without performance degradation.
The paper identifies that synthetic data from stronger models often fails to improve reasoning models due to stylistic divergence, and proposes TESSY, a teacher-student cooperation framework that interleaves token generation to maintain stylistic consistency. On code generation tasks, TESSY improves Qwen3-8B by 11.25% on LiveCodeBench-Pro and 6.68% on OJBench, while teacher-only data causes drops of 3.25% and 10.02%.
A widely adopted strategy for model enhancement is to use synthetic data generated by a stronger model for supervised fine-tuning (SFT). However, for emerging reasoning models like Qwen3-8B, this approach often fails to improve reasoning capabilities and can even lead to a substantial drop in performance. In this work, we identify substantial stylistic divergence between teacher generated data and the distribution of student as a major factor impacting SFT. To bridge this gap, we propose a Teacher-Student Cooperation Data Synthesis framework (TESSY), which interleaves teacher and student models to alternately generate style and non-style tokens. Consequently, TESSY produces synthetic sequences that inherit the advanced reasoning capabilities of the teacher while maintaining stylistic consistency with the distribution of the student. In experiments on code generation using GPT-OSS-120B as the teacher, fine-tuning Qwen3-8B on teacher-generated data leads to performance drops of 3.25% on LiveCodeBench-Pro and 10.02% on OJBench, whereas TESSY achieves improvements of 11.25% and 6.68%.