Researchers have developed WorkFactory, a framework designed to create diverse and executable professional tasks for training AI agents. This system uses information-level planning to build an information support graph, guiding task instantiation and identifying potential failures. The framework has been used to generate over 112,000 tasks across 600 occupations, which were then used to fine-tune Qwen models, resulting in the WorkPilot family of agents. These WorkPilot models demonstrated significant improvements on professional-work benchmarks, with WorkPilot-35B-A3B showing a notable increase in Elo score on GDPval-AA v2.1. AI
IMPACT Introduces a scalable method for generating diverse training data to improve AI agent performance on professional tasks.
RANK_REASON Academic paper detailing a new framework and model training methodology. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
