$ cat wiki/papers/2026/2610.37161-workfactory.md
WorkFactory — executable tasks before expert demonstrations
TL;DR
WorkFactory constructs professional tasks with explicit information dependencies, then probes and repairs their environments before collecting teacher trajectories. The authors report improved professional-work performance after supervised fine-tuning. (source)
Authors & Org
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 and Bryan Dai. Affiliations are not established in the captured abstract record. First submitted 2026-10-27. (source)
Method
Starting from professional intents, an information support graph records dependencies and locations of the information needed for the intended outputs. Probe executions identify support failures; repairs preserve the intended work. Teacher trajectories from the resulting tasks train Qwen models at 4B, 27B and 35B-A3B scales, producing the WorkPilot family. (source)
Results
The authors construct 112,686 tasks across 612 occupations. All three trained models improve over their respective bases on four professional-work benchmarks; WorkPilot-35B-A3B rises from 879 to 1565 Elo on GDPval-AA v2.1. Scaling shows model-dependent diminishing returns, with execution length initially increasing and then decreasing. Component analyses attribute fewer probe-diagnosed environment defects and better downstream performance to planning and repair. (source)
Significance
For Agents (LLM Agents), executable information support becomes part of training-data quality. Unlike simply increasing demonstrations, this method checks whether the intended work can actually be performed. The connection to WebAnything — reconstructing reusable web environments is environment verification; the two studies do not share an evaluation. (source)
Open Questions
The abstract does not establish total data-construction cost, contamination controls or independent reproduction. The authors link released data and models, but the repository and full paper were not inspected. Elo here belongs to GDPval-AA v2.1 and is not comparable with Arena Agent Net Improvement. (source)