WebRetriever Challenge 2026
「 你的 Agent 能在真实网站上完成多少任务？ 」
「 How many real-world web tasks can your Agent complete? 」
「 你的 Agent 能在真实网站上完成多少任务？ 」
「 How many real-world web tasks can your Agent complete? 」
当 AI 智能体走进真实浏览器，它能否像人一样，在充满动态变化的互联网环境中，真正独立完成一项任务？
这是 Web Agent 从实验室走向真实部署的核心瓶颈。长期以来，业界缺一把足够真实的"尺子"——主流 benchmark 多基于少量模拟或自建站点，与真实开放互联网的复杂度存在显著差距；评测维度上，现有方法多关注操作执行的正确性，而对 Agent 能否真正交付最终任务结果缺乏系统性度量。
针对上述不足，明略科技构建了大规模 Web Agent 综合评测基准 WebRetriever，相关论文已被国际顶级学术会议 ECCV 2026 正式接收：
When AI agents enter real browsers, can they truly complete tasks independently in the ever-changing environment of the open internet — just as humans do?
This is the core bottleneck for deploying Web Agents from labs to real business scenarios. For too long, the field has lacked a realistic "yardstick" enough for the open web — most mainstream benchmarks rely on a limited number of simulated or self-hosted sites, falling significantly short of the complexity found on the real open web. In terms of evaluation dimensions, existing methods largely focus on the correctness of individual operations, while lacking systematic measurement of whether an Agent can truly deliver end-to-end task results.
To address these gaps, Mininglamp Technology built WebRetriever, a large-scale comprehensive Web Agent evaluation benchmark. The corresponding paper has been officially accepted by ECCV 2026, a top-tier international academic conference:
横跨科技、金融、医疗、教育、政务等八大领域，全程真实互联网环境。
现有最优方法约 81%，首次让大规模自动评测达到可信精度。
即便最优单一模型，基础导航成功率也不足一半。"到达"远不等于"完成"。
🚀 WebRetriever 挑战赛，诚邀全球学术界和产业界的研究者、开发者与技术团队，在最贴近真实应用的环境中检验 Web Agent 的实战能力——共同推动从"可演示"到"能交付"的关键突破。
Spanning eight domains including technology, finance, healthcare, education, and government — all on the real open internet.
vs. ~81% for the best existing method — enabling large-scale automated evaluation at trustworthy precision for the first time.
Even the top-performing single model achieves less than 50% on basic navigation. "Reaching" a page is far from "completing" a task.
🚀 The WebRetriever Challenge calls upon researchers, developers, and technical teams from academia and industry worldwide to validate the practical performance of Web Agents within an environment that mirrors real-world applications — together driving the pivotal leap from "demo-ready" to "delivery-ready."
挑战赛聚焦 WebRetriever 中最贴近真实部署场景的 Protocol III（端到端任务协议）——要求智能体不仅导航到正确页面，更需从文本、文档、图表等多模态内容中精准抽取目标信息。所有任务遵循三大构造原则：权威性、交互必要性、确定性。
The challenge centers on Protocol III (End-to-End Task Protocol) within WebRetriever, the protocol closest to real deployment scenarios — requiring agents not only to navigate to the correct page, but also to precisely extract target information from text, documents, charts, and other multimodal content. All tasks follow three construction principles: authority, mandatory interaction, and determinism.
具体任务介绍详见备赛期公布的相关文档
Detailed task specifications will be officially released during the preparation period
采用自动化评测框架，参赛团队通过与赛事 Bot 对话发起提测，评测完成后可获得评分反馈。
The competition adopts an automated evaluation framework. Teams initiate evaluation through conversation with the Competition Bot and receive scoring feedback upon completion.
具体分配方案和额外荣誉激励详情随赛事细则公布
Detailed allocation and additional incentives will be announced with competition rules.
备赛期与报名期部分重叠，已注册团队可提前备赛。详细日程在赛事 Octo Space 更新。
Preparation overlaps with registration — registered teams may begin early. Detailed schedule in the Competition Octo Space.
赛事平台 Octo：https://im.deepminer.com.cn/
赛事空间邀请码：0f351ca01bb4c4dd
Competition Platform Octo: https://im.deepminer.com.cn/
Space invite code: 0f351ca01bb4c4dd
个人或团队均可报名，不限国籍、不限机构背景
欢迎学术界、产业界以及独立开发者参与
Open to individuals and teams — no restrictions on nationality or institutional affiliation
Researchers, industry practitioners, and independent developers are all welcome.
明略科技 (2718.HK)
Mininglamp Technology (2718.HK)
