Structure for Reading, Prose for Writing: Asymmetric Structural Conditioning in Multi-Agent Document Authoring

In the world of automated writing, there is a persistent belief that structure is the ultimate key to clarity. Computers have long been taught to read documents more accurately when those documents are organized with clear tags and labels, rather than as a wall of plain text. This principle has become a standard rule in the field of artificial intelligence: if you want a machine to find a specific fact or follow a specific instruction, giving it a structured map of the information helps it navigate the data. Consequently, many engineers have assumed that this same structural advantage applies when the machine is writing. The logic seemed sound: if structure helps a computer read, it should also help a computer write.

However, a new study challenges this assumption by looking at a very specific, high-stakes task: writing formal responses to government tenders. These are complex documents where a company must answer hundreds of questions from a client, often adhering to strict formatting rules and using the client's exact wording. The researchers built a system of multiple artificial intelligence agents to handle this work, separating the tasks of reading the requirements, gathering facts, and drafting the answers. They tested this system against real human-written bids submitted by the same organization. What they discovered was a surprising split in how the machine learns. While structure makes the machine a better reader, it actually makes the machine a worse writer. The study suggests that the best way to guide an AI is not to force it to follow a rigid template, but to let it read structured data and then write in natural language, applying its own internal checks to ensure quality.

The researchers began by deploying their multi-agent system to draft responses for real procurement requests. The system was designed to be cautious and precise. It did not try to guess answers; instead, it broke the work into small steps. First, it read the client's documents and converted them into a structured format that preserved the relationships between questions and instructions. Then, it gathered information from a library of past documents and company records. Finally, it drafted the answers. To see how well it performed, the team compared the machine's drafts against a real bid the organization had written by hand for a different tender. In this comparison, the machine had no example to copy from; it had to rely entirely on its own reasoning and the documents provided.

An independent judge, another artificial intelligence trained to evaluate the quality of writing rather than just how similar it looked to the original, reviewed the results. The machine's answers were rated as at least as good as the human-written ones in forty out of fifty-five sections. In four sections, the machine was actually rated better. It missed no sections entirely. The only major issue was a single unsupported claim, where the machine inferred a security feature that the text did not explicitly state. This result was impressive, especially considering the machine was not a state-of-the-art model and had no prior examples to learn from.

But the researchers did not stop at the surface score. They dug deeper into the fifteen sections where the machine performed worse than the human writer. They wanted to know if the machine failed because it wrote poorly, or because it simply did not have the right information. They found that in nearly seventy percent of these cases, the "failure" was actually a lack of data. The human writer had known certain details from experience or private conversations that the machine's system had never been given. When the researchers adjusted the score to ignore these missing facts, the machine's performance jumped to nearly ninety percent. This revealed a crucial insight: the gap between the machine and the human was not mostly about writing ability, but about information access. The machine was a competent writer that was often handicapped by incomplete briefings.

The study then turned to a more controlled experiment to test the role of structure. The team took the instructions the machine received and presented them in two different ways. In one version, the instructions were written as plain text, like a standard memo. In the other, the instructions were converted into a nested, structured format similar to a digital file with tags. They expected the structured version to help the machine understand the task better, just as it helped with reading. Instead, the opposite happened. When the instructions were structured, the quality of the answers dropped significantly. The machine produced fewer correct answers and repeated the instructions back in its own writing more often, a sign that it was confused.

The researchers realized that structure changes what the machine pays attention to. When the instructions were plain text, the machine focused on the meaning of the request. When the instructions were structured with tags, the machine focused on the tags themselves. It began to treat the client's name or the document title as the most important part of the sentence, leading it to start its answers with awkward phrasing like "The client requires..." instead of answering the question directly. This showed that while structure helps a machine locate information, it can distract a machine when it needs to generate new content.

The team also investigated how the machine handles rules about what not to do. They found that simply telling the machine "do not use this word" often made the problem worse. The machine would fixate on the forbidden word and accidentally include it anyway, or use it in a way that broke the rules. However, when they gave the machine a test to run on its own sentences—asking it to check if a sentence was too similar to the original question and rewrite it if necessary—the errors disappeared. The machine learned to police itself without being told exactly what to avoid.

Finally, the researchers noticed a subtle but dangerous flaw in how the system handled its own notes. The machine would sometimes add a small mark to a document to indicate a specific spot. This mark changed the length of the text slightly. Because the machine processed the text in fixed-size chunks, this tiny change in length shifted the boundaries of the chunks. As a result, the machine would look at a different set of words than intended, leading to a completely different number of questions being extracted. A difference of just two marks in the text caused the system to extract seventeen fewer questions. This showed that even tiny, random changes in how the machine annotates its work could cascade into large errors, simply because the system was too rigid in how it sliced the information.

The overall picture that emerges is one of careful balance. The study does not claim that artificial intelligence has solved the problem of writing formal documents, nor does it suggest that machines are ready to replace human writers entirely. Instead, it offers a precise map of where the technology works and where it struggles. The machine is a powerful tool for reading and organizing information, especially when that information is structured. But when it comes to writing, forcing it into a rigid structure can confuse it. The most effective approach is to let the machine read structured data, but allow it to write in natural language, using self-checks to ensure it stays on track. This distinction between reading and writing is not just a technical detail; it is a fundamental rule for building systems that can be trusted with important tasks. The researchers found that the best results come not from making the machine follow a perfect template, but from giving it the right information and letting it apply its own judgment to the final draft.