AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

UDM-GRPO: Stable and Efficient Group Relative Policy Optimization for Uniform Discrete Diffusion Models

Jiaqi Wang, Haoge Deng, Ting Pan, Yang Liu, Chengyuan Wang, Fan Zhang, Yonggang Qi, Xinlong Wang

ICML 2026 · Reinforcement learning › RL from feedback · approach: RL fine-tuning of generative models · 0 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
use final clean sample as action instead of intermediate prediction x1^t (ablation)URSA + GRPO (backward trajectory) · GenEvaloverall score: 0.84 -> 0.89Improved
use final clean sample as action instead of intermediate prediction x1^t (ablation)URSA + GRPO (backward trajectory) · PickScorePickScore: 21.99 -> 23.10Improved
use final clean sample as action instead of intermediate prediction x1^t (ablation)URSA + GRPO (backward trajectory) · OCRaccuracy: 0.23 -> 0.23No clear effect
train on forward-process trajectory instead of backward trajectory (ablation)URSA + GRPO (final clean sample action) · GenEvaloverall score: 0.89 -> 0.94Improved
train on forward-process trajectory instead of backward trajectory (ablation)URSA + GRPO (final clean sample action) · PickScorePickScore: 23.10 -> 23.51Improved
train on forward-process trajectory instead of backward trajectory (ablation)URSA + GRPO (final clean sample action) · OCRaccuracy: 0.23 -> 0.34Improved
remove classifier-free guidance during training (CFG-Free) (ablation)UDM-GRPO · GenEvaloverall score: 0.94 -> 0.96Improved
remove classifier-free guidance during training (CFG-Free) (ablation)UDM-GRPO · PickScorePickScore: 23.51 -> 23.81Improved
remove classifier-free guidance during training (CFG-Free) (ablation)UDM-GRPO · OCRaccuracy: 0.34 -> 0.57Improved
optimize only three consecutive early high-noise timesteps instead of all or random timesteps (Reduced-Step) (ablation)UDM-GRPO · GenEval and PickScorereward / convergence speedImproved
optimize only three consecutive early high-noise timesteps instead of all or random timesteps (Reduced-Step) (ablation)UDM-GRPO · OCRaccuracyNo clear effect
apply naive Flow-GRPO adaptation (intermediate action, backward trajectory) to URSA (ablation)URSA · long trainingreward stabilityHurt

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