Researchers conducted a systematic analysis of reinforcement learning design space for diffusion models, focusing on likelihood estimation. They found that using an evidence lower bound (ELBO) based model likelihood estimator enables effective, efficient, and stable RL optimization, improving performance on multiple tasks.

A new framework, Ambient Dataloops, refines datasets through an iterative process, improving both the dataset and a diffusion model concurrently. This can lead to better performance in tasks like image generation and protein design.

A new method, probe guidance, is introduced to improve diffusion language models by using frozen internal states to construct a guidance signal, eliminating the need for an additional forward pass at inference time and providing reliable performance on unconditional generation and multiple choice question answering benchmarks.

MUMINS is a new diffusion framework for synthesizing medical image sequences. It uses a single reverse diffusion process to predict anatomical changes and uncertainty maps. The framework is designed to be efficient and reusable across different anatomies via dataset-specific retraining.