WorkFactory: Turning Professional Intents into Executable Work for Scalable Agent Training

What it claims

WorkFactory generates professional work tasks whose required information is present and consistent, so a teacher agent can execute them and produce usable demonstrations. The authors argue that missing or inconsistent task information blocks teacher execution and weakens supervision. They built 112,686 tasks across 612 occupations and fine-tuned Qwen models at 4B, 27B and 35B-A3B on the teacher trajectories, producing the WorkPilot family. All three WorkPilot models beat their own base models on four professional-work benchmarks.

How it works

The pipeline starts from synthesized professional intents. An information-level planning step builds an information support graph. The graph records which pieces of information the intended outputs depend on and where each piece is located in the task environment. The graph then guides task instantiation. Probe executions run on the instantiated task to find support failures, and repairs fix those failures while keeping the intended work unchanged. Teacher trajectories on the finished tasks become supervised fine-tuning data.

flowchart TD
  A[Professional intent] --> B[Information support graph]
  B --> C[Task instantiation]
  C --> D[Probe execution]
  D -->|support failure| E[Repair, keep intended work]
  E --> D
  D -->|coherent| F[Teacher trajectories]
  F --> G[SFT of Qwen bases]

Evidence

WorkPilot-35B-A3B goes from 879 to 1565 Elo on GDPval-AA v2.1, compared with its base model. The abstract does not name the other three benchmarks or give their scores. Scaling the task count shows diminishing returns, and the point where returns flatten depends on the model. Execution length first rises and then falls as training data grows. Controlled component analyses show that planning and repair each reduce task environment defects found by probes, and that they improve downstream performance.

How to use it

The abstract says the data and models are at https://github.com/IQuestLab/WorkFactory. As of October 2026 the Hugging Face page links no models, datasets or Spaces to the paper. The page gives this command to pull the paper into an agent:

hf papers read 2610.37161

Limits

The page has only the abstract and an author post. It does not give the teacher model, the other benchmark names and numbers, training compute, the licence for the data or weights, or the size of the ablation effects. The defect reduction is measured by the same probe executions that drive repair.