What is it about?
We developed a system that breaks legal reasoning into separate, verifiable steps rather than having AI models produce answers in one opaque process. The system first converts legal statutes into structured knowledge representations (formal descriptions of legal concepts and rules), then applies these rules to specific cases through logical reasoning. We tested this on tax calculation problems and found that standard (non-reasoning) AI models improved from 18.8% to 76.4% accuracy when using our structured approach. Importantly, each stage of reasoning, from extracting legal concepts to applying rules to calculating results, can be separately inspected and verified by legal experts.
Featured Image
Photo by Tod S on Unsplash
Why is it important?
In legal applications, understanding how an AI reached its conclusion is often as important as the conclusion itself. Current AI systems typically work as black boxes, making it difficult to verify their reasoning or identify where errors occur. Our approach provides explicit inspection points throughout the reasoning process, allowing legal experts to examine extracted concepts, formalised rules, and inference steps. The research also suggests that structured knowledge representations may help bridge the performance gap between expensive and affordable AI models in domains with well-defined rules.
Perspectives
This work represents an initial investigation focused on calculation-oriented legal reasoning in tax law. The structured approach appears most promising for legal domains with clear rules and numerical outcomes. Further research is needed to test whether similar benefits apply to other areas of law and to evaluate the practical challenges of implementing such systems at scale. The methodology may also be relevant to other rule-governed domains like regulatory compliance and financial analysis, though this remains to be validated.
Albert Sadowski
Warsaw University of Technology
Read the Original
This page is a summary of: On Verifiable Legal Reasoning: A Multi-Agent Framework with Formalized Knowledge Representations, November 2025, ACM (Association for Computing Machinery),
DOI: 10.1145/3746252.3761057.
You can read the full text:
Contributors
The following have contributed to this page
