AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Principled RL for Diffusion LLMs Emerges from a Sequence-Level Perspective

Jingyang Ou, Jiaqi Han, Minkai Xu, Shaoxuan Xu, Jianwen Xie, Stefano Ermon, Yi Wu, Chongxuan Li

ICLR 2026 · Reinforcement learning › RL from feedback · approach: RL for LLM reasoning · 28 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
sequence-level ELBO policy optimization (ESPO) instead of token-level d1/wd1LLaDA-8B-Instruct · Countdownaccuracy: 48.3 -> 81.0 (wd1 vs ESPO avg)Improved
ESPO instead of token-level baselinesLLaDA-8B-Instruct · Sudokuaccuracy: 23.1 -> 86.0 (wd1 vs ESPO avg)Improved
ESPO instead of d1/wd1LLaDA-8B-Instruct · GSM8Kaccuracy: 80.1 -> 82.0 (wd1 vs ESPO avg)Improved
ESPO instead of d1/wd1LLaDA-8B-Instruct · MATHaccuracy: 37.7 -> 39.5 (d1 vs ESPO avg)Improved
ESPO instead of d1/wd1Dream-7B-Instruct · GSM8K / MATHaccuracy: 80.5 -> 81.3 GSM8K; 45.8 -> 46.0 MATH (wd1 vs ESPO)Improved
ESPO instead of d1/wd1LLaDA-8B-Instruct · HumanEval / MBPPpass@1: HumanEval 37.8 -> 40.1; MBPP 37.8 -> 45.4 (vs base LLaDA)Improved
ESPO on Dream Sudoku in 24-token direct-answer settingDream-7B-Instruct · Sudoku (Dream setting)accuracy: 98.1 -> 98.0No clear effect
sequence-level action space with ELBO (vs token-level ELBO or mean-field) (ablation)ESPO · Sudoku training rewardrewardImproved
use k2 KL estimator instead of k1 or k3 (ablation)ESPO · Sudoku training rewardrewardImproved
increase Monte Carlo samples from 1 to 2 or 4 (ablation)ESPO · Sudoku training rewardrewardImproved
increase Monte Carlo samples (ablation)ESPO · Countdown training rewardrewardNo clear effect
vary policy update value mu (8, 12, larger) (ablation)ESPO · Countdown and SudokurewardNo clear effect

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