arXiv 2610.23194
OfficeTown: Generative Organizations for Evolving Agent Intelligence
AI16 upvotes
Authors: Hongchen Wei, Daohong Ying, Zhi Hong, Aidan Wu, Bei Liu, Ruichun Ma, Yifan Yang, Qi Dai +5
Overview
The authors introduce OfficeTown, a framework that generates simulated professional organizations from a natural-language description of their purpose and activities. Members coordinate work, delegate bounded assignments to assistant agents, and review deliverables, while artifacts, decisions, and open matters persist and shape later work. The team converts these work requests, context, and review feedback into OfficeTown Bench and training data.
Introduction
AI agents increasingly need to produce professional deliverables with files and software tools, but real work depends on materials, decisions, and feedback accumulated over time. Existing task setups rarely capture this continuing organizational context.
Results
Supervised fine-tuning and on-policy distillation on OfficeTown data substantially improve Qwen3.5-9B on external professional-work benchmarks, including GDPval. Frontier-model evaluations on OfficeTown Bench (406 tasks, 4,836 evaluation criteria) reveal substantial remaining headroom.
Conclusion
Organizational simulation offers a practical way to generate work-grounded data and environments for developing generalist agents.