arXiv cs.CL· Jihao Zhao, Shuaishuai Zu, Pengfei Li, Biao Qin, Guanghuan Fang, Ben Feng, Baochuan Yang·· 11 小时前AI 评分31
TraceSpec：理解并改进面向 LLM 智能体的投机解码
TraceSpec: Understanding and Improving Speculative Decoding for LLM Agents
AI 导读
TraceSpec 框架将轨迹评估与自适应草稿训练结合，在三种目标模型规模和两种温度下，回放吞吐量比其 DSpark 基线提升 12.5% 至 22.6%。该研究同时发布 TraceSpec-Traj 语料，含 10,000 个训练状态和 2,500 个测试状态，覆盖浏览与软件工程任务。Qwen3-4B 消融显示，自适应分配相比固定整轮监督将接受长度提升 3.3% 至 10.0%。
正文 · 原文
Abstract:LLM agents execute tasks through repeated model calls over evolving histories of reasoning, actions, and tool feedback. Aggregate speculative-decoding metrics merge these distinct decisions, providing limited guidance on where draft supervision is needed. We present TraceSpec, a framework connecting trajectory evaluation with adaptive draft training. It combines dynamic execution with fixed-prefix replay and introduces TraceSpec-Traj, a corpus of 10,000 training states and 2,500 test states spanning browsing and software engineering. Analysis of five speculative baselines reveals that aggregate acceptance can rise with context length while declining within matched late browsing actions. Workload and phase further distinguish these trends, motivating supervision guided by trajectory conditions. TraceSpec trains on intermediate assistant turns and allocates a fixed prediction budget using state-conditioned estimates of prefix-aware acceptance gaps. Across three target sizes and two temperatures, it improves replay throughput by 12.5% to 22.6% over its DSpark base. Qwen3-4B ablations show that adaptive allocation increases accepted length by 3.3% to 10.0% over fixed full-turn supervision, with additional comparisons substantiating the roles of state conditioning and consecutive-token agreement. The resulting evaluation resources, trajectory findings, and training strategy connect the analysis of agent execution to practical draft improvement.
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.41177 [cs.CL] |
| (or arXiv:2610.41177v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.41177 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Jihao Zhao [view email]
[v1]
Fri, 30 Oct 2026 02:57:43 UTC (1,536 KB)
来源：arXiv cs.CL · arxiv.org