WorkFactory: Turning Professional Intents into Executable Work for Scalable Agent Training
WorkPilot-35B-A3B rose from 879 to 1565 Elo on GDPval-AA v2.1 after training on WorkFactory demonstrations.
The researchers built 112,686 tasks spanning 612 occupations. WorkFactory maps the information needed to complete each task, then uses probe executions to find and repair missing or inconsistent support. Teacher demonstrations from those tasks were used to fine-tune three Qwen models, all of which improved over their respective base models across four professional-work benchmarks. Controlled analyses found that both planning and repair reduced task-environment defects and improved downstream performance.
HF Daily Papers' note
The researchers built 112,686 tasks spanning 612 occupations. WorkFactory maps the information needed to complete each task, then uses probe executions to find and repair missing or inconsistent support. Teacher demonstrations from those tasks were used to fine-tune three Qwen models, all of which improved over their respective base models across four professional-work benchmarks. Controlled analyses found that both planning and repair reduced task-environment defects and improved downstream performance.
HF Daily Papers' note
score 5