Diffusion World Models with Heterogeneous Memory Experts

A recent preprint on arXiv presents a framework called Composition of Memory Experts for Diffusion World Models, which separates future-past consistency from a single architecture. This method utilizes a collection of specialized memory experts combined through a contrastive product-of-experts approach. It defines three distinct roles: a short-term memory expert focusing on fine local dynamics, a long-term memory expert that retains episodic history in external diffusion weights through lightweight finetuning at test time, and a third expert designed to address the memory trade-off found in current architectures such as transformers and state-space models. The paper can be accessed at arXiv:2605.18813.