连续扩散语言模型研究实现复兴

连续扩散语言模型（CDLM）在经历数年沉寂后正经历大规模学术复兴1。这一技术范式从2021年早期的离散扩散模型发展而来，2022年开始探索连续方法，但在2023年后因ChatGPT时刻的冲击和训练效率问题（包括Plaid-1B模型效率低64倍）而基本消失，被离散扩散方法主导1。

从2025年下半年开始，CDLM领域出现了转折1。混合方法如CADD、CCDD、CANDI等研究尝试结合离散和连续方法的优势1。到2026年初，包括Categorical Flow Maps、Flow Map Language Models和Discrete Flow Maps在内的三项关键研究相继发表，为连续方法的回归奠定基础1。春季2026年该领域出现"寒武纪爆发"，LangFlow、Spherical flows等多项新研究声称连续扩散与离散扩散已具有竞争力1。

连续扩散方法的核心优势包括更好的步骤蒸馏能力、支持单步采样同时保持令牌间相关性，以及支持基于奖励的微调1。然而，自条件化机制对CDLM性能影响巨大，同时破坏了扩散模型的无状态性假设1。当前该领域面临的主要挑战是评估方法论的完善，GenPPL等指标容易被操纵，已成为CDLM研究的薄弱环节1。

Continuous diffusion language models (CDLMs), a once-prominent approach in AI research, are experiencing a resurgence following a period of dormancy.1 The field has traced a cyclical trajectory: early discrete diffusion methods emerged in 2021, continuous approaches were explored in 2022, but by 2023 continuous methods had largely disappeared as discrete diffusion became the dominant paradigm.1 The shift was driven by factors including the impact of ChatGPT and efficiency concerns, with early continuous models like Plaid-1B demonstrating training efficiency 64 times lower than alternatives.1

The resurgence began in the latter half of 2025 with hybrid methodologies such as CADD, CCDD, and CANDI attempting to combine strengths of both discrete and continuous approaches.1 Early 2026 marked a turning point with three significant research contributions: Categorical Flow Maps, Flow Map Language Models, and Discrete Flow Maps.1 By spring 2026, the field experienced what researchers describe as a "Cambrian explosion" of activity, with multiple studies including LangFlow and Spherical flows claiming that continuous diffusion methods now achieve competitive performance with discrete alternatives.1

The renewed interest is supported by distinct technical advantages of continuous methods. These include superior step distillation capabilities, support for single-step sampling while maintaining inter-token correlations, and compatibility with reward-based fine-tuning.1 However, self-conditioning mechanisms, which significantly impact CDLM performance, undermine the stateless assumptions fundamental to diffusion models.1 A critical weakness in current CDLM research is evaluation methodology; metrics such as GenPPL remain susceptible to manipulation.1