Source: Thomson Reuters

Thomson Reuters explores how legal-focused large language models (LLMs) perform when required to process and analyze long-context legal material such as contracts, regulations, and court opinions. The article presents insights from a new benchmarking initiative focused on the capacity of LLMs to retain, reference, and synthesize information across lengthy legal texts.

The tests involved questions that required detailed reasoning and cross-referencing from multiple sections of documents, mimicking real-world tasks performed by legal professionals. The findings reveal that while current LLMs are improving, they still face notable challenges in accurately handling complex, long-span content. Limitations in memory, context window size, and semantic consistency were observed. Thomson Reuters emphasizes that understanding these limits is crucial for law firms, courts, and in-house legal departments seeking to adopt AI responsibly. The report aims to guide the legal industry toward smarter, more targeted integration of AI tools with real-world legal workflows.