# WorkFactory ### Turning Professional Intents into Executable Work for Scalable Agent Training **⭐ 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 **[IQuest Research](https://github.com/IQuestLab)** [Introduction](#introduction) · [Framework](#framework) · [Performance](#performance) · [Dataset & Models](#dataset) · [Contact](#contact) · [Citation](#citation)

--- ## ✨ News - **[2026.10.27]** 📄 Our paper [WorkFactory](https://arxiv.org/abs/2610.37161) is available on arXiv! ## 💡 Introduction **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. ## 🏭 Framework

WorkFactory framework: information support graphs guide task construction; clean probe executions diagnose blockers and guide repairs before independent teacher trajectory collection.

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](https://arxiv.org/pdf/2610.37161#page=4).* ## 📊 Performance

WorkPilot-35B-A3B performance on GDPval-AA, JobBench, Workspace-Bench, and OmegaUse-OfficeVal, compared with its base model and representative models.

**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](https://arxiv.org/pdf/2610.37161#page=2). WorkPilot Elo scores are local estimates calibrated to GDPval-AA v2.1. See Table 2 and Appendix E for full results and evaluation details.* ## 🤗 Dataset & Models 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 | [![Hugging Face Dataset](assets/badges/huggingface.svg)](#dataset) | | WorkPilot Models | [![Hugging Face Models](assets/badges/huggingface.svg)](#models) | ## 📬 Contact For questions or collaboration, please contact **Ziyang Zeng** ([yling@iquestlab.com](mailto:yling@iquestlab.com)) or **Tingyang Chen** ([chenty@zju.edu.cn](mailto:chenty@zju.edu.cn)). ## 📄 License This repository is licensed under the [Apache License 2.0](LICENSE). ## 📝 Citation If you use WorkFactory or WorkPilot in your research, please cite: ```bibtex @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} } ```