Diffusion Models Boost Agent Repair

Diffusion language models demonstrate significant advantages in local error repair for long-horizon agent planning.

Background: Traditional autoregressive models often regenerate entire plans when environmental changes invalidate assumptions, risking unnecessary alterations and latency. The Plan-and-Patch framework uses parallel unmasking to refill only affected regions while preserving surrounding steps.

Findings: According to a preprint by Syamantak Kumar et al., on the Natural Plan benchmark without task-specific training, DreamReasoner-8B (diffusion) achieved a 53.7% plan repair success rate, nearly double the 27.0% of Qwen3-8B (autoregressive). After task-specific training on ALFWorld and TextCraft, both planners showed similar generation success, but diffusion reduced mean plan-generation latency by 39-46%.

Limitations: Results are author-reported and not yet independently reproduced; validation is primarily on structured program-like planning, with applicability to unstructured complex reasoning remaining uncertain.