AI Paper Atlas

Research map / Reinforcement learning / Offline RL

Noisy-Space Policy Gradient for Diffusion Policies in Offline Reinforcement Learning

Mahmoud Selim, Cristina Cipriani, Karl Henrik Johansson

ICML 2026 · Reinforcement learning › Offline RL · approach: Offline RL sample complexity theory · 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
define value on noisy latents via expected clean-action return (NSAC) instead of applying the clean critic directly to noisy latents (DAC)diffusion actor-critic · AntMaze-large-play-v0normalized score: 50.3 -> 77.8Improved
same change versus DACdiffusion actor-critic · AntMaze-large-diverse-v0normalized score: 55.3 -> 78Improved
same change versus DACdiffusion actor-critic · Pen-Cloned-v1normalized score: 63.9 -> 84.6Improved
same change versus DACdiffusion actor-critic · D4RL locomotion totalnormalized score: 836.4 -> 861.7Improved
same change versus DACdiffusion actor-critic · OGBench visual (4 tasks)aggregate success rate: 255.9 -> 271.5Improved
same change versus DACdiffusion actor-critic · AntMaze totalnormalized score: 275.4 -> 337.2Improved
use excessively strong guidance strength lambda instead of intermediate values {1,3,10} (ablation)NSAC · D4RL taskslearning performanceHurt
increase number of Monte Carlo samples for the noisy-space expectation (ablation)NSAC · D4RL taskslearning performanceNo clear effect
use two Monte Carlo samples instead of more (ablation)NSAC · training costtraining time: approximately 10% faster than prior diffusion actor-criticImproved
vary pessimism coefficient rho over a broad range (ablation)NSAC · D4RL tasksperformanceNo 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