Plan-and-Patch Uses Diffusion Language Models for Agentic Plan Repair
October 9, 2026
Plan-and-Patch employs diffusion language models (dLLMs) to generate and repair structured agent plans through parallel unmasking. This approach allows agents to patch specific failed plan segments while preserving the surrounding prefix and suffix structure.
HOW THIS AFFECTS YOU
●
builderYou can implement more robust long-horizon agents that repair specific steps rather than regenerating entire plans.
●
researcherThis demonstrates the utility of diffusion architectures for non-linear, structured reasoning tasks.