AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Principled RL for Flow Matching Emerges from the Chunk-level Policy Optimization

Yifu Luo, Haoyuan Sun, Xinhao Hu, Penghui Du, Keyu Fan, Bo Li, SiNan Du, Xu Wan, Zhiyu Chen, Bo Xia, Yongzhe Chang, Kai Wu and 3 more

ICML 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants

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
optimize at chunk level instead of step level (GCPO vs Dance-GRPO)GRPO on FLUX.1 Dev · HPDv2.1 testHPSv3: 15.080 -> 15.236Improved
add weighted sampling of chunks by relative L1 distance (ablation)GCPO · HPDv2.1 testHPSv3: 15.236 -> 15.373Improved
add weighted sampling of chunks (ablation)GCPO · GenEvaloverall: 0.69 -> 0.67Hurt
add weighted sampling of chunks (ablation)GCPO · DPGoverall: 86.60 -> 85.14Hurt
use temporal-dynamics adaptive chunking instead of fixed chunk size (ablation)GCPO · HPDv2.1 testHPSv3: 15.173 -> 15.236Improved
fixed chunking [4, 4, 4, 4] versus step-level GRPO (ablation)GCPO (fixed) · HPDv2.1 testHPSv3: 15.080 -> 15.078No clear effect
replace outcome reward with a process reward (ablation)Dance-GRPO · HPDv2.1 testHPSv2: 0.365 -> 0.348Hurt
switch reward model from HPSv3 to PickScore (ablation)GCPO · HPDv2.1 testHPSv3: 14.612 -> 14.810 (vs Dance-GRPO)Improved
use L2 distance instead of relative L1 to define chunks (ablation)GCPO · HPDv2.1 testHPSv3: 15.236 -> 15.268No clear effect
use cosine distance instead of relative L1 to define chunks (ablation)GCPO · HPDv2.1 testHPSv3: 15.236 -> 15.220No clear effect
treat the whole trajectory as a single chunk (no temporal dynamics) (ablation)GCPO · HPDv2.1 testHPSv3: 15.142 vs 15.236 (fixed [16] vs adaptive)Hurt
train on high-noise chunks only (ablation)GCPO · HPDv2.1 testHPSv3 improvementMixed

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