Researchers have developed an AI system that can author formal documents by reading a requester's forms and writing against them. They tested this system on a real-world tender-response task and found that its performance was comparable to human-written bids in many cases. However, they also discovered that the system performed worse when it had to condition on structured data rather than prose instructions. This led the authors to propose an 'asymmetric structural conditioning' approach, where structured markup is used for reading tasks but not for writing tasks. They demonstrated this approach with several experiments and found improved performance in some cases. --- Why it matters: This research matters because it highlights the importance of tailoring AI systems to specific tasks and data formats. By understanding how different types of input affect an AI's performance, developers can create more effective and efficient systems that better serve their intended purpose. Source: https://arxiv.org/abs/2608.20786