---
id: 20261120-T0-06
title: "投机解码新发现：草稿模型隐藏层宽度不必对齐目标模型"
title_en: "NarrowSpec: Draft Hidden Width Need Not Match the Target in Speculative Decoding"
url: https://ai.daily.yangsir.net/daily/20261120-T0-06
issue_date: 2026-11-20
publish_date: 2026-11-19T05:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2611.23835
---

# 投机解码新发现：草稿模型隐藏层宽度不必对齐目标模型

投机解码通过小模型起草、大模型验证来加速大语言模型生成，而读取目标模型隐藏状态的草稿模型（如DFlash）通常被认为需要与目标模型保持一致的隐藏层宽度。研究提出NarrowSpec，证明草稿模型的隐藏宽度可以窄于目标模型，仍能有效利用目标模型的隐藏状态完成起草。这一发现降低了此类草稿模型的设计约束和计算开销，为投机解码的加速方案提供了更灵活的结构选择。

## English Version

**NarrowSpec: Draft Hidden Width Need Not Match the Target in Speculative Decoding**

Speculative decoding speeds up LLM generation by having a small draft model propose tokens that the target model verifies in one forward pass. Draft models that read the target's hidden states, such as DFlash, were assumed to need matching hidden width. NarrowSpec shows that the draft's hidden width can be narrower than the target's while still effectively leveraging target hidden states. The result relaxes a key design constraint and lowers compute overhead for this class of draft models, giving speculative decoding more flexible architectural options.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2611.23835)

**详情页**：https://ai.daily.yangsir.net/daily/20261120-T0-06

---

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