Research map / Reinforcement learning / RL from feedback
Slow-Fast Policy Optimization: Reposition-Before-Update for LLM Reasoning
Ziyan Wang, Zheng Wang, Xingwei Qu, Qi Cheng, Jie Fu, Shengpu Tang, Minjia Zhang, Xiaoming Huo
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 |
|---|---|---|---|
| replace GRPO with SFPO | GRPO on Qwen2.5-Math-1.5B · Math benchmarks avg | average accuracy: 38.35 -> 40.19 | Improved |
| replace GRPO with SFPO | GRPO on DS-distilled-Qwen-1.5B · Math benchmarks avg | average accuracy: 47.73 -> 50.53 | Improved |
| replace GRPO with SFPO | GRPO on DS-distilled-Qwen-1.5B · AIME25 | accuracy: 23.33 -> 30.83 | Improved |
| replace GRPO with SFPO | GRPO on DS-distilled-Qwen-7B · Math benchmarks avg | average accuracy: 60.47 -> 63.04 | Improved |
| replace GRPO with SFPO | GRPO on Qwen2.5-Math-7B · Math benchmarks avg | average accuracy: 48.36 -> 49.19 | Improved |
| apply SFPO on top of DAPO | DAPO · Math benchmarks avg | average accuracy: 49.30 -> 50.56 (DS-1.5B), 46.48 -> 47.87 (Qwen3-4B-Base) | Improved |
| apply SFPO instead of GRPO to reach same accuracy | GRPO · Efficiency | rollouts required: 3.21x, 3.50x, 4.93x fewer rollouts | Improved |
| apply SFPO instead of GRPO for same accuracy | GRPO · Efficiency | training time: 2.62x, 2.65x, 4.19x speedups | Improved |
| apply SFPO instead of GRPO on coding | GRPO · LiveCodeBench | Pass@1 (Avg 4) peak: 18.91 -> 21.77 | Improved |
| add Stage II interpolation (reposition) with alpha<1 (ablation) | SFPO · Training curves, K=7 | training accuracy | Improved |
| add slow correction (Stage III) (ablation) | SFPO · Training curves | training accuracy | Improved |
| add entropy control (alpha set to 0 after z-score trigger) (ablation) | SFPO · DS-Qwen-7B training curves | accuracy | Improved |
Closest papers by meaning
- WS-GRPO: Weakly-Supervised Group-Relative Policy Optimization for Rollout-Efficient Reasoning (ICML 2026 · similarity 0.92)
- Accelerating RL for LLM Reasoning with Optimal Advantage Regression (NeurIPS 2025 · similarity 0.92)
- One-Way Policy Optimization for Self-Evolving LLMs (ICML 2026 · similarity 0.91)
- Group-in-Group Policy Optimization for LLM Agent Training (NeurIPS 2025 · similarity 0.91)
- Act Only When It Pays: Efficient Reinforcement Learning for LLM Reasoning via Selective Rollouts (NeurIPS 2025 · similarity 0.91)
- Sample More to Think Less: Group Filtered Policy Optimization for Concise Reasoning (ICLR 2026 · 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.
