AI Paper Atlas

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

ICLR 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants · 5 citations (OpenAlex)

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 changedWhereMeasuredEffect
replace GRPO with SFPOGRPO on Qwen2.5-Math-1.5B · Math benchmarks avgaverage accuracy: 38.35 -> 40.19Improved
replace GRPO with SFPOGRPO on DS-distilled-Qwen-1.5B · Math benchmarks avgaverage accuracy: 47.73 -> 50.53Improved
replace GRPO with SFPOGRPO on DS-distilled-Qwen-1.5B · AIME25accuracy: 23.33 -> 30.83Improved
replace GRPO with SFPOGRPO on DS-distilled-Qwen-7B · Math benchmarks avgaverage accuracy: 60.47 -> 63.04Improved
replace GRPO with SFPOGRPO on Qwen2.5-Math-7B · Math benchmarks avgaverage accuracy: 48.36 -> 49.19Improved
apply SFPO on top of DAPODAPO · Math benchmarks avgaverage accuracy: 49.30 -> 50.56 (DS-1.5B), 46.48 -> 47.87 (Qwen3-4B-Base)Improved
apply SFPO instead of GRPO to reach same accuracyGRPO · Efficiencyrollouts required: 3.21x, 3.50x, 4.93x fewer rolloutsImproved
apply SFPO instead of GRPO for same accuracyGRPO · Efficiencytraining time: 2.62x, 2.65x, 4.19x speedupsImproved
apply SFPO instead of GRPO on codingGRPO · LiveCodeBenchPass@1 (Avg 4) peak: 18.91 -> 21.77Improved
add Stage II interpolation (reposition) with alpha<1 (ablation)SFPO · Training curves, K=7training accuracyImproved
add slow correction (Stage III) (ablation)SFPO · Training curvestraining accuracyImproved
add entropy control (alpha set to 0 after z-score trigger) (ablation)SFPO · DS-Qwen-7B training curvesaccuracyImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy