---
id: 20261103-T0-10
title: "TraceSpec：拆解LLM智能体的投机解码，找出草稿该在哪里加监督"
title_en: "TraceSpec: Breaking Down Speculative Decoding for LLM Agents to Target Draft Supervision"
url: https://ai.daily.yangsir.net/daily/20261103-T0-10
issue_date: 2026-11-03
publish_date: 2026-11-02T05:00:00.000Z
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2610.41177
---

# TraceSpec：拆解LLM智能体的投机解码，找出草稿该在哪里加监督

论文研究LLM智能体场景下的投机解码。智能体通过反复调用模型、在推理与行动、工具反馈不断累积的历史上执行任务，而现有的聚合式投机解码指标把这些不同决策混在一起，难以指导草稿模型该在哪些环节加强监督。TraceSpec给出更细粒度的分析框架，用于定位需要草稿监督的具体决策点。若能落地，意味着智能体推理加速不必全局加码，而是把算力花在真正拖慢速度的环节上。

## English Version

**TraceSpec: Breaking Down Speculative Decoding for LLM Agents to Target Draft Supervision**

The paper examines speculative decoding for LLM agents, which execute tasks through repeated model calls over evolving histories of reasoning, actions, and tool feedback. Aggregate speculative-decoding metrics merge these distinct decisions, offering limited guidance on where draft supervision is needed. TraceSpec provides a finer-grained analysis framework to locate the specific decision points that require draft supervision. If it translates into practice, agent inference speedups could come from targeting the steps that actually slow things down rather than applying uniform effort.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2610.41177)

**详情页**：https://ai.daily.yangsir.net/daily/20261103-T0-10

---

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