Brief, mostly facts

OfficeTown generates persistent agent organizations for training data

The framework builds simulated organizations and ships OfficeTown Bench with 406 tasks across 23 organizations.

What the research found

OfficeTown is a generative simulation framework that builds persistent professional organizations to produce agent training data, with OfficeTown Bench of 406 tasks and 4,836 criteria across 23 organizations.

Claims and how well the evidence supports them

OfficeTown Bench has 406 tasks and 4,836 evaluation criteria across 23 simulated organizations 99%
Supervised fine-tuning on OfficeTown data improves Qwen3.5-9B on GDPval 96%
OfficeTown spans eight professional domains 98%

Evidence

huggingface.co https://huggingface.co/papers/2610.23194
arxiv.org https://arxiv.org/abs/2610.23194

Actionability rated low.