Harvey, an AI-driven platform designed to streamline legal workflows, has overhauled its review tables algorithm, delivering more accurate answers, granular citations, and faster results. The updates aim to save legal teams significant time when reviewing massive datasets such as contracts, regulatory filings, or trial exhibits.

Review tables are a key feature of Harvey, enabling users to transform thousands of documents into structured grids that can be scanned, filtered, and acted upon. However, inefficiencies in the previous algorithm—such as redundant fields and less precise citation methods—had created friction in workflows. The revamped approach addresses these issues head-on.

Sharper Answers, Improved Reasoning

The updated algorithm now consolidates what were previously two separate outputs—summary and additional context—into a single, unified answer. Alongside it, a new reasoning field breaks down how conclusions are reached, offering transparency into the model’s decision-making process. This change is particularly useful for legal use cases requiring interpretation, as users can now assess not just the outcome but the logic behind it.

For example, when the tool is asked to determine whether a supply agreement includes a change-of-control provision, the “reasoning” field will outline where the algorithm searched and how it concluded that the provision was absent. This level of detail isn’t just about accuracy—it’s about empowering lawyers to validate results quickly and make informed decisions with confidence.

Granular Citations and Speed Boost

Another significant enhancement involves moving from cell-level to sentence-level citations. Previously, citations were attached to entire cells, which often made it difficult to trace specific statements back to their source. The new sentence-based methodology ensures each assertion in a response can be directly tied to a precise location in the underlying document, streamlining the verification process for time-pressed legal professionals.

This granular approach also boosted output speed. By rethinking its execution pipeline, Harvey achieved lower latency at scale, even for complex datasets. For instance, a table with 30 columns and 1,000 documents generates 30,000 cells—yet the new algorithm maintains quick response times across this workload.

Evaluation Results: A Clear Win

Harvey rigorously tested the new algorithm through side-by-side preference evaluations conducted by contract attorneys and Applied Legal Researchers (ALRs). The results were striking: the updated system was preferred four times more often than the original, with even stronger preference in complex scenarios like credit agreements or trial exhibits. Analytical depth and transparency in reasoning emerged as the top drivers of user satisfaction.

Equally important, the enhancements held up under benchmarks for reliability and latency. Combined with improvements in citation accuracy, the overall update represents a meaningful leap forward for legal professionals who rely on Harvey for document review.

What’s Ahead

The new algorithm lays the groundwork for future developments. Harvey plans to introduce dynamic rows and columns that can derive directly from multiple knowledge sources, including entire datasets and web pages. This next phase will bring multi-source citations and confidence-based scoring into the platform, further enhancing its utility for high-stakes legal work.

For legal teams under pressure to deliver precise, defensible results at speed, these updates signal Harvey’s commitment to evolving its tools in a way that aligns with the rigorous demands of the profession.

Image source: Shutterstock