WebAnything: From the Wild Web to Scalable Agent Environments
Computer Science > Artificial Intelligence arXiv:2610.37786 (cs) [Submitted on 27 Oct 2026] Title:WebAnything: From the Wild Web to Scalable Agent Environments Authors:Tianci Xie, Zikun Cai, Chaoyun Zhang, Pinxue Guo, Bingxuan Li, Rui Zheng, Hang Yan View a PDF of the paper titled WebAnything: From the Wild Web to Scalable Agent Environments, by Tianci Xie and 6 other authors View PDF HTML (experimental)
计算机科学 > 人工智能 arXiv:2610.37786 (cs) [提交于 2026 年 10 月 27 日] 标题：WebAnything：从真实网络到可扩展的智能体环境 作者：Tianci Xie, Zikun Cai, Chaoyun Zhang, Pinxue Guo, Bingxuan Li, Rui Zheng, Hang Yan 查看标题为《WebAnything：从真实网络到可扩展的智能体环境》的论文 PDF，作者为 Tianci Xie 及其他 6 位作者 查看 PDF HTML（实验性）
Abstract:Web agents require diverse, executable, and resettable website environments for reproducible training and evaluation. Human-assisted mirrors demand per-site engineering, while generative construction can diverge from specified real websites. Capture-and-replay preserves observations, with interaction coverage constrained by captured content and replay capabilities. We present WebAnything, a fully automated framework that reconstructs source-grounded, reusable Web environments from real URLs without per-site manual implementation or repair after initial setup and authorization.
摘要：Web 智能体需要多样化、可执行且可重置的网站环境，以进行可重复的训练和评估。人工辅助的镜像需要针对每个站点进行工程设计，而生成式构建可能偏离指定的真实网站。捕获并重放技术虽然保留了观测结果，但交互覆盖范围受限于已捕获的内容和重放能力。我们提出了 WebAnything，这是一个全自动框架，它能从真实 URL 重构出基于源的、可重用的 Web 环境，且在初始设置和授权后无需针对每个站点进行手动实现或修复。
An Explorer-Builder collects browser-observable structures, resources, and interaction outcomes; an independent verifier checks frozen candidates and guides local repair or targeted source recapture. We construct 244 website artifacts across 14 application domains and 53 categories, supporting 12,984 executed and accepted tasks. Compared with human-assisted WebHarbor, WebAnything achieves higher aggregate scores on three fidelity metrics, with 12 wins, two ties, and no losses in a blind audit.
一个“探索者-构建者”（Explorer-Builder）模块负责收集浏览器可观测的结构、资源和交互结果；一个独立的验证器负责检查冻结的候选环境，并指导局部修复或针对性的源数据重新捕获。我们构建了涵盖 14 个应用领域和 53 个类别的 244 个网站工件，支持 12,984 项已执行并被接受的任务。与人工辅助的 WebHarbor 相比，WebAnything 在三项保真度指标上取得了更高的总分，在盲审中取得了 12 胜、2 平、0 负的成绩。
Fine-tuning Qwen3.5-9B on 5,856 interaction trajectories improves REAL success by 9.9 percentage points, Odysseys-Easy Rubric Avg and Task Avg by 6.25 and 6.32 percentage points, and WebArena-Verified Overall success by 13.33 percentage points under the fixed harness. Code: this https URL Comments: Subjects: Artificial Intelligence (cs.AI) Cite as: arXiv:2610.37786 [cs.AI] (or arXiv:2610.37786v1 [cs.AI] for this version)
在 5,856 条交互轨迹上对 Qwen3.5-9B 进行微调，使 REAL 成功率提高了 9.9 个百分点，Odysseys-Easy 评分平均值和任务平均值分别提高了 6.25 和 6.32 个百分点，在固定测试框架下，WebArena 验证的总体成功率提高了 13.33 个百分点。代码：此 HTTPS URL 评论：学科：人工智能 (cs.AI) 引用方式：arXiv:2610.37786 [cs.AI]（或此版本的 arXiv:2610.37786v1 [cs.AI]）