Sorting the noise: a field guide to diffusion in RL

What it does This repository is a curated reading list tracking research that applies diffusion models to reinforcement learning. It catalogs papers by venue and year, grouping work that uses diffusion for trajectory planning, policy optimization, data synthesis, and world models. A brief overview at the top outlines the two dominant paradigms—planning by iterative trajectory denoising and using conditional diffusion as an expressive policy class—and lists claimed benefits over traditional RL.

The interesting bit The list organizes the field around two foundational papers: Diffuser, which frames planning as a diffusion probabilistic model, and Diffusion-QL, which treats the policy itself as a conditional diffusion model. That split is useful because it clarifies why generative image-modeling machinery is being imported into control problems in two fundamentally different ways.

Key highlights

• Papers are grouped by conference—NeurIPS, ICML, ICLR, CVPR, ICRA—and arXiv, with tags for key ideas and experiment environments like D4RL, MuJoCo, and Robomimic.
• A short “Advantage” section claims diffusion RL bypasses bootstrapping for long-term credit assignment, avoids discounting-induced myopia, and scales easily to multi-modal data.
• The maintainers state the list is continually updated to track the “frontier of Diffusion RL.”
• Includes a “Codebase” section in the table of contents for associated implementations.

Caveats

• The synthesis is thin: beyond a two-paragraph overview and a three-bullet advantage list, the repository is essentially a formatted bibliography with minimal commentary.
• Several sections list future conference years (e.g., ICML 2026, ICLR 2026), suggesting the timeline mixes preprints with speculative or forthcoming venue placement.

Verdict Worth bookmarking if you are actively working on generative policies or trajectory planning and need a quick index of what to read. Skip it if you are looking for a unified library or framework to run experiments—this is strictly a reading list.