arXiv:2511.19473v2 Announce Type: replace 
Abstract: Diffusion Language Models (DLMs) have shown strong potential for text generation and are becoming a competitive alternative to autoregressive models. The denoising strategy plays an important role in determining the quality of their outputs. Mainstream denoising strategies include Standard Diffusion and BlockDiffusion. Standard Diffusion performs global denoising without restricting the update range, often finalizing incomplete context and causing premature end-of-sequence predictions. BlockDiffusion updates fixed-size blocks in a preset order, but its rigid structure can break apart coherent semantic units and disrupt reasoning. We present WavefrontDiffusion, a dynamic decoding approach that expands a wavefront of active tokens outward from finalized positions. This adaptive process follows the natural flow of semantic structure while keeping computational cost equal to block-based methods. Across four benchmarks in reasoning and code generation, WavefrontDiffusion achieves state-of-the-art performance while producing outputs with higher semantic fidelity, showing the value of adaptive scheduling for more coherent and efficient generation.

يقدم WavefrontDiffusion جدولًا زمنيًا ديناميكيًا للتشفير لنماذج اللغة القائمة على الانتشار (DLM)، مما يعزز توليد النص من خلال توسيع جبهة من الرموز النشطة من المواقع النهائية. تتناول هذه الطريقة قيود الأساليب الحالية مثل الانتشار القياسي وBlockDiffusion، والتي يمكن أن تعيق التفكير والتماسك في النص الناتج.

WavefrontDiffusion presenta un calendario de decodificación dinámica para los Modelos de Lenguaje por Difusión (DLM), mejorando la generación de texto al expandir un frente de ondas de tokens activos desde posiciones finalizadas. Este enfoque aborda las limitaciones de los métodos existentes como la Difusión Estándar y la BlockDiffusion, que pueden obstaculizar el razonamiento y la coherencia en el texto generado.

WavefrontDiffusion présente un calendrier de décodage dynamique pour les modèles de langage par diffusion (DLM), améliorant la génération de texte en élargissant un front d'ondes de tokens actifs à partir de positions finalisées. Cette approche répond aux limitations des méthodes existantes telles que la diffusion standard et la BlockDiffusion, qui peuvent entraver le raisonnement et la cohérence dans le texte généré.

WavefrontDiffusion introduces a dynamic decoding schedule for Diffusion Language Models (DLMs), enhancing text generation by expanding a wavefront of active tokens from finalized positions. This approach addresses limitations of existing methods like Standard Diffusion and BlockDiffusion, which can hinder reasoning and coherence in generated text.

WavefrontDiffusion: Dynamic Decoding Schedule for Improved Reasoning

Was this article worth reading? Share it

LucidQuery AI

Airparser

LangWatch

Dubsmart LLC

Dynamiq

All Voice Lab

Ready to build your own newsroom?