Registration for the WebRetriever Global Challenge is officially open!

BEIJING, July 28, 2026 /PRNewswire/ -- This competition is hosted by Mininglamp Technology and co-organized by Peking University, the Institute of Automation of the Chinese Academy of Sciences (CASIA), Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences, and Machine Heart.

Featuring a total prize pool of $15,000, the competition is open to both individuals and teams, with no restrictions on nationality or institutional background. We warmly invite participants from academia, industry, and independent developer communities to check the registration details below and compete alongside top talent from around the world!

Background

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:

• Industry-Leading Scale

Covers 800 live websites and 1,550 cross-industry tasks across eight domains including technology, finance, healthcare, education, and government services—all run on the real open internet.

• Accuracy

The self-developed NavEval framework achieves 91.2% agreement with human expert judgments (vs. ~81% for the best existing method), enabling large-scale automated evaluation at trustworthy precision for the first time.

• Revealing the Gap

Evaluation data shows that even the top-performing single model achieves less than 50% on basic navigation success, while end-to-end task completion rate is only ~20%. "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, we drive the pivotal leap from "demo-ready" to "delivery-ready."

Task

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.

• Participating Agents are required to:
  • Receive a natural language task instruction
  • Autonomously devise and execute full action pipelines on live websites (page navigation, element interaction, form filling, multi-step transitions, etc.)
  • Extract and return the result information required by the task

• Task characteristics:
  • End-to-end evaluation tasks based on the WebRetriever dataset, covering real website scenarios across multiple industries
  • Tasks are designed based on dynamic pages of live websites, involving high-level challenges such as multi-step reasoning and cross-page information aggregation

• Detailed task specifications will be officially released during the preparation period.

Evaluation

The competition adopts an automated evaluation framework. Teams initiate evaluation through conversation with the Competition Bot and receive scoring feedback upon completion.

Multiple submissions are allowed within each evaluation window; the best score counts as the stage result.

All participating agents must connect to the unified evaluation environment via OpenAI-compatible APIs. Teams are fully responsible for their own model inference resource management.

Detailed evaluation specifications will be officially released during the preparation period.

Prizes

• Total prize pool: $15,000 USD.
• Detailed allocation and additional incentives will be announced with competition rules.

Schedule

Registration (3 weeks): 07/16 — 08/07
Register, learn about the task background, and complete team sign-up.

Preparation (4 weeks): Late July — Late August
Evaluation guidelines released; teams debug Agents and connect to the test environment.

The preparation period partially overlaps with registration—registered teams may begin preparation early.

Submission (1 week): Late August
Submission window opens; automated evaluation and scoring.

Review: Early September
Score verification and final ranking confirmation.

Results Announcement: September
Final leaderboard published; winners awarded.

Detailed schedule updates will be posted in the dedicated WebRetriever Competition Octo Space.

Register

• Competition Platform Octo：https://im.deepminer.com.cn/
• Space invite code: 0f351ca01bb4c4dd

Step 1 — Create an Octo account (skip if you already have one)

• Option A: Sign up at the link above. We recommend registering with your email; if you use another method, please bind your email in account settings afterward.
• Option B (recommended for developers): If you have Claude Code, ChatGPT Codex, Cursor, or any AI coding assistant, you can complete registration entirely from the terminal — no browser interaction needed. See the https://mininglamp-ai.github.io/WebRetriever_Challenge/join/

Step 2 — Join the Competition Space

Log in and use the invite code above to join the WebRetriever Competition Space

Step 3 — Complete Registration

Once inside the Space, follow the instructions to submit your registration details (team name, members, etc.).

Eligibility

Individuals or teams may register. No restrictions on nationality or institutional affiliation. Researchers, industry practitioners, and independent developers are all welcome.

Organizers & Support

• Organizer: Mininglamp Technology
• Co-organizers:
  • Peking University;
  • Institute of Automation, Chinese Academy of Sciences (CASIA);
  • Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences;
  • Machine Heart

• Organizing Committee：

Mininglamp Technology (2718.HK)
Wei Dong — Mano model design & development, WebRetriever co-first author
Tianyu Fu — Mano model algorithm lead, WebRetriever co-first author
Zhe Yu — Director of Data Products, Mano data pipeline & evaluation research
Yuting Liao — WebRetriever dataset curation & competition operations
Han Lin — Datasets research, WebRetriever dataset curation
Chenxu Zhao — Vice President, Chief Multimodal Scientist, WebRetriever corresponding author
Minghui Wu — Founder, CEO & CTO, Project Leader

• Academic Advisors：

Zhen Lei — Ph.D., IEEE / IAPR / AAIA Fellow. Researcher at the Institute of Automation, CAS; Professor at the University of Chinese Academy of Sciences; Professor at the Centre for Artificial Intelligence and Robotics (CAIR), Hong Kong Institute of Science & Innovation, CAS; Doctoral supervisor. Research: video analysis and understanding, multimodal large models, biometric recognition. 200+ papers, 41,000+ citations, H-index 92. Listed in the Global Top 2% Scientists Ranking.

Ping Wang — Ph.D., Professor & doctoral supervisor at Peking University; State Council Special Allowance Expert. Director of the Intelligent Computing & Sensing Lab, National Engineering Research Center for Software Engineering. MOE Higher Education Scientific Research Output Award (Natural Science), 1st Prize, first contributor. Research: intelligent computing & sensing, smart healthcare, information security. 200+ papers; 40+ patents and software copyrights.

Meng Ma — Ph.D., Associate Researcher at Peking University. Deputy Director of the National Engineering Research Center for Software Engineering; Research Lead of the Intelligent Computing and Sensing Lab. Member of multiple CCF technical committees; Member of CCF YOCSEF (Young Computer Scientists & Engineers Forum). Research: intelligent network O&M, fault diagnosis and prediction, situational awareness. 80+ papers, 2,700+ citations.

Kun Yan — Ph.D., Distinguished Associate Researcher at PKU School of Computer Science. Selected for National Postdoctoral Researcher Funding Program; CCF Digital Medicine Division Executive Committee. Research: low-annotation learning, 3D visual segmentation, smart healthcare. 21 papers in T-PAMI, CVPR, AAAI, ACM MM, Nature Communications. ACM MM 2024 Best Paper Nomination.

Contact Us: WebRetrieverChallenge@mininglamp.com