On-Policy Residual Learning for Diffusion Language Models

Recent advancements in Diffusion Language Models include ORL, a framework for on-policy residual learning that utilizes model-induced residual trajectories and entropy-contraction to improve speed-quality Pareto frontiers. Additionally, research embeds DLM dynamics within open quantum systems, introducing a tau-leaping unraveling scheme for generation conditioned on superposed prompts. Meanwhile, Routing Policy Optimization (RPO) enhances online diffusion reinforcement learning by transforming attention routing into stochastic exploration actions, reducing costly rollout-based optimization while preserving the original sampler.