Plan-and-Patch boosts plan repair with diffusion language models

The paper introduces Plan-and-Patch, a framework where a diffusion language model (dLLM) creates a structured plan and then fixes only the broken parts. In tests on the Natural Plan benchmark, the diffusion planner repaired plans 53.7% of the time, compared with 27.0% for an autoregressive planner. After training on agentic tasks like ALFWorld, both approaches generated plans at similar success rates, but the diffusion method was 39‑46% faster. The work suggests a way to make long‑horizon agents more efficient and adaptable.