---
id: 20260626-T0-08
title: "Dustfin：解决投机解码验证瓶颈，大幅提升长文本生成效率"
title_en: "Dustin: Solving Speculative Decoding Bottleneck for Efficient Long-Context Generation"
url: https://ai.daily.yangsir.net/daily/20260626-T0-08
issue_date: 2026-06-26
publish_date: 2026-06-25T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2606.24957
---

# Dustfin：解决投机解码验证瓶颈，大幅提升长文本生成效率

针对大模型长文本生成中投机解码（Speculative Decoding）面临的验证瓶颈问题，新研究提出了Dustin方法。现有的加速方案往往受限于键值（KV）缓存加载带来的延迟，导致多批次处理效率低下。Dustin通过草稿增强的稀疏验证技术，减少了对KV缓存的依赖，从而在不牺牲生成质量的前提下显著提高了推理吞吐量。这为处理长文档和高并发请求提供了一种更高效的解码策略。

## English Version

**Dustin: Solving Speculative Decoding Bottleneck for Efficient Long-Context Generation**

Addressing the verification bottleneck in speculative decoding for long-context LLMs, new research introduces Dustin. Existing acceleration methods are often limited by Key-Value (KV) cache loading latency during multi-batch processing. By utilizing draft-augmented sparse verification, Dustin reduces dependency on KV cache access, significantly improving inference throughput without sacrificing generation quality.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2606.24957)

**详情页**：https://ai.daily.yangsir.net/daily/20260626-T0-08

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*