AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Simple Policy Gradients for Reasoning with Diffusion Language Models

Anthony Zhan

ICML 2026 · Reinforcement learning › RL from feedback · approach: RL for LLM reasoning · 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 stepwise denoising-MDP policy gradient (AGRPO) instead of ELBO-based sequence-level objective (SPG)LLaDA-8B-Instruct · Countdown (len 512)accuracy (%): 70.3 -> 82.4Improved
use AGRPO instead of SPGLLaDA-8B-Instruct · GSM8K (len 512)accuracy (%): 84.5 -> 88.1Improved
use AGRPO instead of SPGLLaDA-8B-Instruct · MATH-500 (len 256)accuracy (%): 40.0 -> 39.4Hurt
use AGRPO instead of SPGLLaDA-8B-Instruct · MATH-500 (len 512)accuracy (%): 41.8 -> 42.0No clear effect
use AGRPO instead of SPGLLaDA-8B-Instruct · Sudoku (len 512)accuracy (%): 93.1 -> 95.2Improved
post-train with AGRPO then sample with few steps (m=32)LLaDA-8B-Instruct · GSM8K test, n=384accuracy (%): 59.7% at 12 tokens per stepImproved
post-train with AGRPO to reduce sampling stepsLLaDA-8B-Instruct · GSM8K test, n=384sampling steps for equal accuracy: 4x fewer stepsImproved
increase MC timestep samples k from 2 to 32 (ablation)AGRPO · GSM8K trainingconvergence speed / wall time: 44.8s -> 72.4s per stepMixed
increase k from 32 to 128 (ablation)AGRPO · GSM8K trainingconvergence speed / wall time: 72.4s -> 158.1s per stepNo clear effect
add low-discrepancy and entropy importance sampling (ablation)AGRPO · Countdown, n=128, m=64reward convergence and gradient normImproved
add KL penalty (beta>0) (ablation)AGRPO · reasoning tasksaccuracyNo clear effect
use LoRA fine-tuned AGRPO vs GRPO on AR modelAGRPO · GSM8K (len 512)accuracy gain over base: +9.9% vs +5.3% for GRPOImproved

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