⭐ Star us if you find this useful!
Ziyang Zeng · Tingyang Chen · Jia Deng · Qiyao Wang
Shuo Tang · Yike Yang · Tao Ren · Wayne Xin Zhao
Feng Chang · Jujie He · Chuan Hao · Ran Tao · Bryan Dai
Introduction · Framework · Performance · Dataset & Models · Contact · Citation
- [2026.10.27] 📄 Our paper WorkFactory is available on arXiv!
WorkFactory turns professional intents into executable tasks for training working agents. It uses information-level planning to organize the facts needed for each deliverable, then execution-based repair to fix missing or inconsistent inputs while preserving the intended work.
We construct 112,686 tasks across 612 occupations and train the WorkPilot family at 4B, 27B, and 35B-A3B scales. All three models improve over their base models across four professional-work benchmarks.
WorkFactory builds professional intents from occupations, skills, and real deliverables. An information support graph guides workspace construction, and clean probe executions check whether the task can be completed. Retained environments are used to collect independent teacher demonstrations for supervised fine-tuning.
Figure 2 from the paper.
WorkPilot-35B-A3B improves from 879 to 1565 Elo on GDPval-AA v2.1, alongside gains on JobBench, Workspace-Bench, and OmegaUse-OfficeVal. WorkPilot-27B reaches 1612 Elo, while WorkPilot-4B reaches 1375 Elo.
Figure 1 from the paper. WorkPilot Elo scores are local estimates calibrated to GDPval-AA v2.1. See Table 2 and Appendix E for full results and evaluation details.
The full training corpus covers 112,686 tasks, 612 occupations, and 23 occupational groups, with an average of 4.15 reference files per task. Teacher demonstrations train WorkPilot-4B, WorkPilot-27B, and WorkPilot-35B-A3B.
| Resource | Hugging Face |
|---|---|
| WorkFactory Dataset | |
| WorkPilot Models | |
For questions or collaboration, please contact Ziyang Zeng (yling@iquestlab.com) or Tingyang Chen (chenty@zju.edu.cn).
This repository is licensed under the Apache License 2.0.
If you use WorkFactory or WorkPilot in your research, please cite:
@misc{zeng2026workfactory,
title = {WorkFactory: Turning Professional Intents into Executable Work for Scalable Agent Training},
author = {Ziyang Zeng and Tingyang Chen and Jia Deng and Qiyao Wang and Shuo Tang and Yike Yang and Tao Ren and Wayne Xin Zhao and Feng Chang and Jujie He and Chuan Hao and Ran Tao and Bryan Dai},
year = {2026},
eprint = {2610.37161},
archivePrefix = {arXiv},
primaryClass = {cs.AI},
url = {https://arxiv.org/abs/2610.37161}
}