投机草稿与连续扩散语言模型的生成能力差距
原标题：The Generative Gap Between Speculative Drafter and Continuous Diffusion Language Model
AI 摘要
该论文比较了连续扩散语言模型（CDLM）与投机解码草稿模型在无验证、独立生成条件下的生成能力。作者用 LangFlow 实例化 CDLM，用 DSpark 的 Markov 与循环头架构实例化草稿模型，在 OpenWebText 上从零训练并评估 GenPPL 与 MAUVE。结果显示在长度 128 和 1024 时 LangFlow 两项指标均优于 DSpark 变体，而 DSpark-RNN 仅在长度 16 领先；循环记忆可改善草稿模型但无法弥合长序列差距，增加 LangFlow 去噪步数可进一步提升指标。
正文节选
Yujian Chen, Hao Jiang, Weijian Luo, Guang Cheng, Tianyang Hu The Generative Gap Between Speculative Drafter and Continuous Diffusion Language Model Abstract Continuous diffusion language models (CDLMs) and speculative drafting offer two approaches to efficient language generation, but their generative capabilities are typically evaluated under different conditions. CDLMs generate independently through iterative refinement, whereas speculative drafters produce proposals that a target model verif