ContractScrub is a new benchmark designed to evaluate how well Large Language Models (LLMs) perform "contract scrubbing"—the final, meticulous review of legal agreements for errors and inconsistencies. While LLMs are frequently used for legal text processing, this research identifies a gap in how these models perform on the routine, high-stakes task of identifying drafting defects before a contract is finalized.
The Task of Contract Scrubbing
Contract scrubbing is a repetitive but essential legal process where practitioners check for issues like misused defined terms, incorrect cross-references, and inconsistent language. These errors are often subtle and embedded in long documents. The authors, including researchers from Thomson Reuters and Imperial College London, argue that while this task aligns with the general capabilities of frontier LLMs—such as long-context reasoning and named entity recognition—there has been no formal evaluation of how these models perform in a real-world legal setting.
How the Benchmark Works
The researchers constructed the ContractScrub dataset using 44 contracts sourced from the CUAD dataset. To ensure high-quality, realistic testing, the benchmark was developed through a three-stage pipeline: 1. Selection: Contracts were screened for structural integrity and length (typically 10–15 pages). 2. Annotation: Experienced lawyers with at least 8 years of practice identified existing drafting errors. 3. Augmentation: The lawyers purposefully inserted additional, realistic drafting errors to create a "gold standard" for evaluation.
The final dataset contains 3,014 annotated tasks across nine specific categories, such as "Undefined Capitalized Terms" and "Incorrect Section References." Models are evaluated on their ability to identify these defects, with a focus on recall, as missing a legal error is generally more costly than flagging a potential false positive.
Performance Results
The study evaluated nine frontier and open-weight models, including versions of GPT, Claude, Qwen, and Gemini. The results indicate that contract scrubbing is significantly more difficult for LLMs than their performance on general-purpose benchmarks might suggest.
The best-performing model achieved a macro-average recall of 0.750, and all models recorded F1 scores below 0.650. Performance was inconsistent across categories: models were relatively proficient at identifying explicit lexical signals, such as defined terms, but struggled significantly with categories requiring the inference of intent within a specific context, such as "Incorrect Capitalization in Context."
Why Domain-Specific Benchmarking Matters
The authors conclude that general-purpose legal benchmarks often test narrow skills in isolation, failing to capture the complexity of reviewing an entire, long-form document. The findings suggest that even advanced models fall short of the performance expected based on their general capabilities when applied to specialized, high-precision tasks. This highlights the necessity of using narrowly targeted, domain-specific benchmarks to measure the real-world impact of AI in professional fields like law.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!