首页 > AI前沿 > Fast-dLLM++: Fr\'echet Profile Decoding for Faster Diffusion LLM Inference

Fast-dLLM++: Fr\'echet Profile Decoding for Faster Diffusion LLM Inference

arXiv自然语言 2026-06-02 07:18 14 阅读 查看原文
用户现在需要把给定的新闻正文排版成语义化的HTML片段,规则要遵守。首先,先看正文的结构,开头是一段背景,然后是提出的问题,然后是Fast-dLLM++的方案,然后是实验结果,最后是代码地址。首先,要不要加小标题?用户说长文用h2或h3划分小节,那可以分几个部分对吧? 等下,先理清楚段落:首先第一段是讲Diffusion LLM的瓶颈,Fast-dLLM的不足,对吧?然后接下来是提出Fast-dLLM++的方法,然后是实验结果,最后是代码地址?不对,先仔细拆分自然段落: 哦原文的自然段落:第一个自然段是“Diffusion large language models promise parallel token generation, yet inference remains bottlenecked by deciding which masked tokens can be safely committed together. Fast-dLLM addressed this with KV caching and confidence-guided parallel decoding, but its decoding theory uses a homogeneous high-confidence assumption that effectively reduces each candidate set to its weakest selected token.” 然后下一句是“We argue that this leaves speed on the table because real decoding steps exhibit heterogeneous confidence profiles.” 不对,是不是原文的段落?哦用户给的正文是连在一起的?不对,我再看用户给的正文: 正文是: Diffusion large language models promise parallel token generation, yet inference remains bottlenecked by deciding which masked tokens can be safely committed together. Fast-dLLM addressed this with KV caching and confidence-guided parallel decoding, but its decoding theory uses a homogeneous high-confidence assumption that effectively reduces each candidate set to its weakest selected token. We argue that this leaves speed on the table because real decoding steps exhibit heterogeneous confidence profiles. We propose Fast-dLLM++, a training-free extension that introduces Fréchet profile decoding: selecting parallel commit sets from the full sorted confidence profile rather than a single worst-case confidence. The resulting rule is a heterogeneous-confidence generalization of Fast-dLLM's factor selector and it recovers the previous rule exactly in the equal-confidence case and adds a provable heterogeneity bonus when the selected tokens have uneven confidences. Fast-dLLM++ leaves the model, diffusion process, and cache implementation entirely unchanged, making it a drop-in replacement for existing Fast-dLLM decoding. Experiments on GSM8K, MATH, HumanEval, and MBPP with the LLaDA-8B model show that the theoretical improvement translates directly into empirical gains: profile-aware selection improves the accuracy-throughput frontier by exploiting safe parallelism that weakest-token rules miss, achieving up to 37% higher throughput at comparable accuracy. Our code release is at https://github.com/Ringo-Star/FastdLLM_plusplus. 哦,那如果要划分小节的话,可以分“问题背景”

?不对,或者

研究背景与现有方法局限

,然后下面的p放前面的内容,然后

提出的改进方案

,然后p放方案的内容,然后

实验验证与成果

,然后p放实验的内容,然后最后一句是代码地址,是不是可以放最后?不对,要保留所有信息,不能增删。 等下,规则说每个自然段落用p包裹,长文用h2/h3划分小节。那首先,划分小节是合理的对吧?首先:

研究背景与现有方法局限

Diffusion large language models promise parallel token generation, yet inference remains bottlenecked by deciding which masked tokens can be safely committed together. Fast-dLLM addressed this with KV caching and confidence-guided parallel decoding, but its decoding theory uses a homogeneous high-confidence assumption that effectively reduces each candidate set to its weakest selected token. We argue that this leaves speed on the table because real decoding steps exhibit heterogeneous confidence profiles.

然后

改进方案

We propose Fast-dLLM++, a training-free extension that introduces Fréchet profile decoding: selecting parallel commit sets from the full sorted confidence profile rather than a single worst-case confidence. The resulting rule is a heterogeneous-confidence generalization of Fast-dLLM's factor selector and it recovers the previous rule exactly in the equal-confidence case and adds a provable heterogeneity bonus when the selected tokens have uneven confidences. Fast-dLLM++ leaves the model, diffusion process, and cache implementation entirely unchanged, making it a drop-in replacement for existing Fast-dLLM decoding.

然后

实验结果与资源开放

Experiments on GSM8K, MATH, HumanEval, and MBPP with the LLaDA-8B model show that the theoretical improvement translates directly into empirical gains: profile-aware selection improves the accuracy-throughput frontier by exploiting safe parallelism that weakest-token rules miss, achieving up to 37% higher throughput at comparable accuracy. Our code release is at https://github.com/Ringo-Star/FastdLLM_plusplus.

等下,是不是可以?有没有增删内容?没有,小标题是划分小节的,不算增内容?不对,等下规则说“保留原文所有信息,不得增删改内容”,哦,小标题是我加的?不对,会不会有问题?哦不对,规则说“长文用

/

划分小节”,那小标题是对内容的概括,是不是可以?不对,等下,有没有可能用户说的划分小节是指原文的结构,如果没有原文的标题,那是不是可以不用加?不对,再仔细看规则:“2.每个自然段落用

包裹,长文用

/

划分小节”。哦,那如果原文的内容可以划分为几个部分,加小标题是合理的?不对,等下,会不会违反“不得增删改内容”?哦,不对,小标题是结构标识,不是正文内容?不对,再想想,有没有可能用户的意思是,如果原文有小标题就保留,没有的话,要不要加?不对,再看用户给的正文是纯文本的新闻正文,没有现成的小标题。那我是不是可以划分小节,用h2,内容是对小节的概括,但是不改变原文内容。 或者,也可以不用加小标题?不对,规则说长文用h2/h3划分,这个正文不算短,所以应该划分。 等下,再看,重要句子可以用strong,比如“up to 37% higher throughput”是重要的,所以加strong。 然后,有没有需要列表或者引用或者代码的地方?没有,代码这里就是链接,不用pre,所以不用。 等下,再检查内容有没有遗漏:原文的所有句子都放进去了吗? 第一部分的p:Diffusion...heterogeneous confidence profiles. 对,都在里面。 第二部分的p:We propose...Fast-dLLM decoding. 对,都在里面。 第三部分的p:Experiments...FastdLLM_plusplus. 对,都在里面。 小标题的话,是不是合适?比如

研究背景与现有局限

,

Fast-dLLM++核心方案

,

实验验证与代码开放

,这样也可以。 等下,会不会有人觉得小标题是添加的内容?不对,规则说可以用h2/h3划分小节,所以划分小节的小标题是允许的,因为是结构排版,不是正文内容的增删。 哦,不对,再仔细看规则4:“保留原文所有信息,不得增删改内容”,这里的增删改是指正文内容,排版用的结构标签和小标题是为了语义化排版,是规则要求的,所以没问题。 或者,有没有可能我不用加小标题,直接把整个作为几个p?不对,规则说长文用h2/h3划分,所以应该加。 等下,再调整一下,让小标题更准确:

研究背景与现有方法局限

Diffusion large language models promise parallel token generation, yet inference remains bottlenecked by deciding which masked tokens can be safely committed together. Fast-dLLM addressed this with KV caching and confidence-guided parallel decoding, but its decoding theory uses a homogeneous high-confidence assumption that effectively reduces each candidate set to its weakest selected token. We argue that this leaves speed on the table because real decoding steps exhibit heterogeneous confidence profiles.

Fast-dLLM++核心改进方案

We propose Fast-dLLM++, a training-free extension that introduces Fréchet profile decoding: selecting parallel commit sets from the full sorted confidence profile rather than a single worst-case confidence. The resulting rule is a heterogeneous-confidence generalization of Fast-dLLM's factor selector and it recovers