Contributors

Haoyu Zhao1,*Sitian Tu1,*Yuxuan Huang2,*Huichi Zhou1Zhangyue Yin1Jiahao Huang3Jingyuan Xu3Tianmiao Shi1Jin Shi1Nianxi Zhang1Yangmiao Zhang4Zhuolang Chan1Huiyan Hu1Chenchao Sheng1Yaqing Luo3Yue Chen5Huaxiu Yao6Meng Fang2Jun Wang1,†
1 University College London2 University of Liverpool3 Imperial College London4 University of Electronic Science and Technology of China5 Peking University6 University of North Carolina at Chapel Hill

* Equal contribution. † Corresponding author: jun.wang@cs.ucl.ac.uk

Domain Expertise & Task Review

Domain experts author professional assignments and define the criteria used to evaluate them. Each admitted task undergoes automated review and independent review by two human reviewers.

Figure 1.How a task enters CLE, following Figure 4 of the manuscript.
  1. Expert DesignA domain expert drafts the task from a workflow in their own practice: inputs, deliverable and checklist for every level.
  2. Agent Review & Real RunA review agent audits the draft with a human-authored review skill, then runs it end to end. Verdict: Strong Accept, Accept, Borderline, Major Revision or Reject.
  3. Dual Human ReviewTasks rated Accept or higher go to two randomly assigned reviewers, who review independently. Both must accept.
  4. Admitted to CLEThe 120 tasks that cleared both gates form the benchmark.
Feedback: a rejection at either gate returns the task to its author for revision.

The author list and affiliations above follow the current manuscript. Additional contributor acknowledgements will be added when the list is confirmed.

Contact

For questions about the research, contact the corresponding author, Jun Wang.
jun.wang@cs.ucl.ac.uk ↗