Recursive Sparse Reasoning Enhances Multimodal Diffusion Models

A new study has unveiled a recursive, sparse mixture-of-experts framework designed to improve structured reasoning in multimodal text-to-image diffusion models. This approach is inspired by how humans think in modules and integrates a recursive aspect within joint attention layers. It progressively enhances visual tokens through multiple latent steps while efficiently sharing parameters by selectively using certain neural modules. A gating network is responsible for picking the specialized modules at each stage. This research addresses the challenge of applying latent reasoning and recursion from language models to continuous visual tokens in text-to-image generation. You can find the complete findings on arXiv under the identifier 2604.25299.