Planning is becoming increasingly vital for long‑horizon agents, where successful execution requires coordinating sub‑goals, tool usage, and intermediate outcomes across many steps. However, assumptions made during planning can be invalidated by the environment, tools may return unexpected results, and actions can fail. Effective agents therefore need not only to generate plans but also to revise them. Revisions usually affect only a local region of a plan, leaving the surrounding structure intact. Instead of regenerating the entire plan and risking unnecessary changes, repair can regenerate the affected region conditioned on the preserved prefix and suffix. We introduce the Plan-and-Patch framework: a diffusion language model (dLLM) generates a structured, program‑like plan via parallel unmasking and repairs it by filling selected regions while keeping surrounding steps fixed. We compare DreamReasoner‑8B (diffusion) and Qwen3‑8B (autoregressive, AR) as planners. On the Natural Plan benchmark without task‑specific training, diffusion achieves a plan‑repair success rate of 53.7%, nearly twice that of AR (27.0%). After task‑specific training on agentic benchmarks such as ALFWorld and TextCraft, both planners attain similar plan‑generation success, but diffusion reduces mean plan‑generation latency by 39%–46% relative to AR. These results demonstrate that Plan-and-Patch provides a faster plan‑generation and effective plan‑repair framework for long‑horizon agents.
Review