[论文] Online Skill Learning for Web Agents via State-Grounded Dynamic Retrie...

论文概要

研究领域: ML
作者: Jiaxi Li, Ke Deng, Yun Wang
发布时间: 2025-06-01
arXiv: 2606.04391

中文摘要

语言智能体日益依赖可复用技能来改进跨相关任务的多步网页自动化。越来越多的工作研究在线技能学习，智能体从先前任务轨迹中不断归纳技能并即时复用于未来任务。然而，现有方法主要在任务级别复用技能：基于初始任务指令检索一组固定技能，并在整个执行过程中保持不变。这种静态策略与网页执行不一致，其中适当的下一步动作不仅取决于任务目标，还取决于当前网页状态，而后者常转变为初始技能无法覆盖的情境。为解决这一差距，我们提出状态 grounded 动态检索（SGDR），一种在线技能学习方法，使网页智能体能够逐步复用技能。SGDR包含三个组件：滑动窗口提取过程，将完成的轨迹转化为可在中间执行状态调用的可复用子程序；双重文本-代码表示，连接技能检索与可执行动作；状态 grounded 动态检索机制，将技能与任务目标和当前网页状态同时匹配。在WebArena跨五个领域的实验显示，SGDR持续优于强基线，使用GPT-4.1和Qwen3-4B分别实现37.5%和24.3%的平均成功率，相对于最强基线分别提升10.6%和10.0%。代码可在此https URL获取。

原文摘要

Language agents increasingly rely on reusable skills to improve multi-step web automation across related tasks. A growing line of work studies online skill learning, where agents continually induce skills from previous task trajectories and reuse them in future tasks on the fly. However, existing methods mainly reuse skills at the task-level: a fixed set of skills is retrieved based on the initial task instruction and then held fixed throughout execution. This static strategy is misaligned with web execution, where the appropriate next action depends not only on the task goal but also on the current webpage state, which often transitions into situations that the initial skills fail to cover. To address this gap, we propose State-Grounded Dynamic Retrieval (SGDR), an online skill learning m...

自动采集于 2026-06-05

#论文 #arXiv #ML #小凯