Plan-and-Patch: Diffusion LLM Repairs Agent Plans Twice as Often as Qwen3-8B

The diffusion language model DreamReasoner-8B repairs agent plans twice as often as the autoregressive Qwen3-8B: 53.7% versus 27.0% on the Natural Plan benchmark, without task-specific fine-tuning. This is the main result of the Plan-and-Patch preprint on arXiv, which proposes a planning scheme for long-horizon agents. After dedicated training on ALFWorld and TextCraft, both models plan with similar success, but the diffusion model produces a plan 39–46% faster.

How Plan-and-Patch works

The scheme belongs to the plan-and-act class. A planner first builds a structured, program-like plan, and an executor then follows its steps. In long scenarios, a plan rarely survives contact with reality: assumptions go stale, a tool returns an unexpected response, an action fails. An agent has to be able to revise a plan, not just make one.

The usual approach for an autoregressive model is to generate the plan from scratch. That is slow and risky: the model may change steps that were fine for no reason. The authors propose a different approach. Only the affected section of the plan is masked, while the beginning (prefix) and the end (suffix) stay fixed. The diffusion model fills in the gap using both edges as context. It also builds the original plan through parallel unmasking, that is, not token by token from left to right.

My reading is that bidirectional context suits repair best. Fixing a step in the middle depends both on what came before it and on where the plan has to end up. An autoregressive model has to carry that condition through the prompt, while a diffusion model has it available directly in its generation mechanism. The authors point to this mechanism, but the abstract does not prove a causal link.

What the measurements showed

The authors compared two models of the same scale, 8 billion parameters each: DreamReasoner-8B as the diffusion planner and Qwen3-8B as the autoregressive one. The repair test ran without training for the specific task.

• Natural Plan, plan repair without fine-tuning: 53.7% for the diffusion model versus 27.0% for the autoregressive one, a gap of almost two times.
• ALFWorld and TextCraft after fine-tuning: plan generation success is similar for both models.
• Average plan generation latency: 39–46% lower for the diffusion model.
• Unit of editing: a selected region of the plan, with the surrounding steps unchanged.

The picture is mixed, which adds to its credibility. Diffusion does not make plans better where models are trained for the benchmark: generation success evens out. Its advantage shows up in repair without specialization and in speed.

Limitations and practical takeaways

This is still a preprint without peer review, and all the data comes from 8B-class models and text environments. The abstract does not make clear how the scheme would behave on larger models, with real tools, or with noisy APIs. ALFWorld and TextCraft remain simulations, and failures in production are messier.

The work offers three takeaways for people building agents. First, store the plan as a structure with stable steps, so it can be repaired locally rather than only regenerated. Second, evaluate repair separately from generation, because models with the same success on the initial plan can differ sharply in repair success (here, 53.7% versus 27.0%). Third, count latency for every replanning call. An agent in a long scenario replans many times, and a 39–46% gain on a single call accumulates over the course of the task.

If the results hold up on other models and environments, diffusion language models will get a practical niche: a compact planner that quickly produces a structure and edits it surgically, leaving the heavy tool work to the executor.