先思考，快速扩散：通过自回归计划条件改善扩散语言模型推理

Diffusion large language models (dLLMs) generate text through iterative denoising but consistently underperform in multi-step reasoning tasks. This study hypothesizes that the performance gap arises from a coordination issue: autoregressive (AR) models build coherence token-by-token, while diffusion models must coordinate all positions simultaneously. A method called plan conditioning is proposed, which prepends a short (~100-token) natural-language plan from an AR model to the prompt of the diffusion model. This plan acts as a frozen scaffold, providing globally visible context that every token position can attend to from the first denoising step. On the GSM8K dataset, plan conditioning improves the performance of LLaDA-8B-Instruct from 75.6% to 87.2% (+11.6 percentage points), matching the performance of a same-size AR model (LLaMA 3.1 8B, 87.7%).