Plan-and-Patch: Diffusion Language Models for Agentic Planning

2026-10-07

AI Summary

This study addresses the inefficiency of traditional autoregressive models in long-horizon agent plan generation and local repair, where they often compromise overall structural integrity. To overcome these limitations, this work proposes the Plan-and-Patch framework, which leverages diffusion language models (dLLMs) alongside programmatic plan representations to enable parallel generation of structured plans. Furthermore, it introduces a conditional masked decoding technique that performs localized infilling within specific regions while preserving preceding and subsequent steps, thereby circumventing the need for global regeneration. Experimental results demonstrate that the proposed framework improves plan repair success rates to 53.7%, nearly tripling the performance of autoregressive baselines. Additionally, post-task-training latency is reduced by 39–46%, significantly enhancing overall agent execution efficiency.

Abstract

Planning is increasingly important for long-horizon agents, where successful execution requires coordinating subgoals, tool use, and intermediate outcomes over many steps. Yet assumptions made during planning may be invalidated by the environment, tools may return unexpected results, or actions may fail. Effective agents must therefore not only generate plans, but also revise them. Such revisions often affect only part of a plan, leaving the preceding and subsequent structure intact. Rather than regenerate the entire plan and risk unnecessary changes, repair can regenerate the affected region conditioned on the preserved prefix and suffix. We introduce Plan-and-Patch, a plan-and-act framework in which a diffusion language model (dLLM) generates a structured, program-like plan through parallel unmasking and repairs it by filling in selected regions while keeping the surrounding steps fixed. We compare DreamReasoner-8B and Qwen3-8B as diffusion and autoregressive (AR) planners. On Natural Plan without task-specific training, diffusion (53.7%) achieves nearly twice the plan repair success rate of AR (27.0%). After task-specific training on agentic benchmarks, ALFWorld and TextCraft, the planners achieve similar observed success in plan generation, while diffusion reduces mean plan-generation latency by 39-46% relative to AR. Our results show that Plan-and-Patch provides a framework for faster plan generation and effective plan repair in long-horizon agents.

Problem

• Agentic Planning
• Plan Repair
• Long-horizon Agents
• Diffusion Language Models

Innovation

• Diffusion Language Models
• Agentic Planning
• Plan Repair
• Parallel Unmasking
• Long-horizon Agents