AI RESEARCH
MetaState: Persistent Working Memory Enhances Reasoning in Discrete Diffusion Language Models
arXiv CS.AI
•
ArXi:2603.01331v2 Announce Type: replace-cross Discrete diffusion language models (dLLMs) generate text by iteratively denoising a masked sequence. However, standard dLLMs condition each denoising step solely on the current hard-masked sequence, while intermediate continuous representations are discarded after sampling and remasking. We term this bottleneck the \textbf{Information Island} issue: continuous information remains isolated within individual denoising steps and fails to propagate across the trajectory.