Research map / Large language models / Instruction and fine-tuning
SRFT: A Single-Stage Method with Supervised and Reinforcement Fine-Tuning for Reasoning
Yuqian Fu, Tinghong Chen, Jiajun Chai, Xihuai Wang, Songjun Tu, Guojun Yin, Wei Lin, Qichao Zhang, Yuanheng Zhu, Dongbin Zhao
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 |
|---|---|---|---|
| Add entropy-aware SFT weight w_SFT (instead of fixed 1.0) (ablation) | SRFT · Math reasoning avg (5 benchmarks) | avg accuracy: 55.1 -> 59.1 | Improved |
| Add entropy-aware RL weight w_RL (instead of fixed 1.0) (ablation) | SRFT · Math reasoning avg (5 benchmarks) | avg accuracy: 56.2 -> 59.1 | Improved |
| Apply RL then SFT instead of SFT then RL | Qwen2.5-Math-7B · Math reasoning avg (5 benchmarks) | avg accuracy: 52.5 -> 37.4 | Hurt |
| Add KL constraint to RL->SFT | RL->SFT · Math reasoning avg (5 benchmarks) | avg accuracy: 37.4 -> 38.3 | No clear effect |
| Run SFT then RL instead of RL alone | Qwen2.5-Math-7B · Math reasoning avg (5 benchmarks) | avg accuracy: 49.4 -> 52.5 | Improved |
| Run SFT then RL instead of SFT alone | Qwen2.5-Math-7B · Math reasoning avg (5 benchmarks) | avg accuracy: 47.3 -> 52.5 | Improved |
| Use single-stage SRFT instead of two-stage SFT->RL | Qwen2.5-Math-7B · Math reasoning avg (5 benchmarks) | avg accuracy: 54.5 -> 59.5 | Improved |
| Use single-stage SRFT instead of best SFT+RL baseline on OOD | Qwen2.5-Math-7B · OOD avg | avg accuracy: 57.8 -> 62.5 | Improved |
| Use SRFT instead of the best zero-RL baseline | Qwen2.5-Math-7B · Math reasoning avg (5 benchmarks) | avg accuracy: +9.0 points | Improved |
| Use SRFT instead of SFT | Qwen2.5-Math-7B · Math reasoning avg (5 benchmarks) | avg accuracy: +4.8 points | Improved |
| Use SRFT instead of SFT+RL methods | Qwen2.5-Math-7B · Math reasoning avg (5 benchmarks) | avg accuracy: +3.4 points | Improved |
| Use the SRFT chat template instead of the Qwen2.5-Math template (ablation) | Qwen2.5-Math-7B base · Math reasoning avg | Average: 0.218 -> 0.340 | Improved |
Later papers that compared with this method
1 later accepted paper reports SRFT as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Beyond Two-Stage Training: Cooperative SFT and RL for LLM Reasoning (ICML 2026 · better on 3 of 3 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
- UFT: Unifying Supervised and Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.93)
- RL Squeezes, SFT Expands: A Comparative Study of Reasoning LLMs (ICLR 2026 · similarity 0.92)
- AceReason-Nemotron 1.1: Advancing Math and Code Reasoning through SFT and RL Synergy (ICLR 2026 · similarity 0.92)
- Reasoning with Exploration: An Entropy Perspective (AAAI 2026 · similarity 0.91)
- BREAD: Branched Rollouts from Expert Anchors Bridge SFT & RL for Reasoning (NeurIPS 2025 · similarity 0.91)
- Diversity-Aware Policy Optimization for Large Language Model Reasoning (NeurIPS 2025 · similarity 0.91)
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.
