[论文] ReTrace: Distilling Tool-Using Agents into Planner-Free Workflows

论文概要

研究领域: ML 作者: Kaiyuan Hou, Yang Ping Kuo, Indranil Bhattacharya, Rohith Nama, Arushi Gupta, Carlos D. Castillo, Prathap Ramachandra, Xiaofan Jiang, Francesco Carbone 发布时间: 2026-11-10 arXiv: 2611.14492

中文摘要

业务流程自动化对新的输入一遍又一遍地运行同一套标准操作规程（SOP）。LLM 智能体可以执行自然语言写成的 SOP，但每一步都要调用 LLM 来选择，规模扩大后延迟与成本不可忽视。我们提出 ReTrace：把一份 SOP 与几段记录的智能体运行转化为可复用工作流——选择下一步的是代码而非 LLM。智能体在每次工具调用时写下的推理往往说明了为什么采取某动作，光凭工具调用本身则不行。ReTrace 提取每一步背后的条件，跨运行合并为紧凑规则文档，再从中生成可执行工作流。只有最后这阶段依赖平台，规则文档可编译到不同编排平台。在我们生产平台 79 套流程上，ReTrace 工作流完成率 93.8%（智能体 96.8%），单任务成本低 24%。在 τ²-Bench 上任务成功率 75.8%（智能体 80.6%），运行 1.3 秒而非 170 秒；对训练失败三轮修复可把 k=2 工作流从 66.7% 提升到 81.2%。SOP-Bench 8 套流程上，无修复的 ReTrace 略逊少样本生成（80.6% vs 83.6%），从带标注失败中修复后达 87.8%、单任务 0.40 秒，对比智能体 85.5%、46.6 秒。

原文摘要

Business process automation runs the same Standard Operating Procedure (SOP) again and again on new inputs. Large Language Model (LLM) agents can follow an SOP written in natural language, but they call an LLM to choose every step of every run, which adds latency and cost at scale. We present ReTrace, which turns an SOP and a few logged agent runs into a reusable workflow in which code, not an LLM, selects the next step. The agent's reasoning, the explanation it writes with each tool call, often states why it took an action; the tool calls alone do not. ReTrace extracts the condition behind each step, merges these conditions across runs into a compact rule document, and generates an executable workflow from it. Only this last stage depends on the platform, so rule documents can be compiled...

#论文 #arXiv #ML #小凯