Converting legislative documents into structured XML is a high-stakes task. A single dropped citation or a mislabeled amendment can propagate into the legal texts that lawyers and the public rely on. The source document must be reproduced exactly, preserving section hierarchy, verbatim wording, and the precise spans of amendments.
A new arXiv preprint evaluates how well different models handle this task and whether a "strong harness"—likely a structured prompting or decoding framework—can close the gap between weaker and stronger models. According to the abstract, the harness narrows the model gap, suggesting that careful engineering around the model matters as much as model choice.
The announcement is brief, so details on the specific models, dataset, and evaluation metrics are not yet available. But the central finding is clear: a strong harness can meaningfully reduce errors in legislative XML conversion, which is critical for downstream legal applications.