Zach Anderson
Mar 11, 2026 20:21
Harvey AI unveils BLB: Research benchmark targeting hard agentic legal research problems that current foundation models still can’t solve reliably.
Legal AI company Harvey has released BigLaw Bench: Research, a new benchmark designed to expose where frontier AI models still fail at complex legal research tasks—even when equipped with web search tools.
The benchmark, developed with data partner Snorkel AI, focuses on U.S. case law research problems that leading foundation models currently cannot solve reliably. Harvey’s goal is straightforward: find the breaking points where AI stops being useful to practicing lawyers.
Why Search-Based Benchmarking Matters Now
BLB: Research marks Harvey’s first end-to-end benchmark requiring models to use search tools, identify relevant context, and deliver cited responses. The shift reflects an industry-wide consensus that search has become the primary method for grounded AI responses.
Models are increasingly trained to use search rather than rely on static training data for current knowledge. Testing them without search tools, Harvey argues, undersells their actual capabilities. At the same time, law firm clients expect pin-cited sources as standard for research-driven work.
Finding Where Models Actually Break
The benchmark deliberately targets realistic complexity rather than obscure edge cases. Harvey identified three task categories that matter to practicing lawyers: finding the best case to cite for a legal proposition, drafting research memos, and planning claims or defenses.
Through testing, Harvey found that model answers become unhelpful once they complete less than 60% of required task criteria. At that threshold, models typically miss critical reasoning steps, take wrong research turns, or deliver analysis too shallow to be actionable.
Sample tasks span 14 practice areas including corporate law, securities litigation, privacy, intellectual property, and tax. One example involves assessing earn-out manipulation claims after a Delaware asset sale where the buyer consolidated operations and reassigned key staff. Another evaluates securities fraud claims against an EV company whose CEO allegedly misrepresented prototype functionality before a 45% stock decline.
Broader Implications for Legal AI
The release comes amid significant momentum in foundation model development. Yann LeCun’s new World Model AI lab raised $1 billion in Europe’s largest seed round on March 10, while France’s Living Models secured $7 million for specialized foundation models on March 11.
Harvey positions BLB: Research as infrastructure for testing improvements across three vectors: raw model capabilities, unique data sources, and the tooling connecting them. The same search capabilities tested here also underpin AI-powered searches of SEC filings, investigation documents, deal rooms, and internal firm knowledge bases.
For legal tech investors and law firms evaluating AI vendors, the benchmark offers a concrete framework for assessing whether a model can actually accelerate legal research or just generate plausible-sounding noise. The 60% completion threshold provides a useful mental model: below that line, you’re probably better off doing the research yourself.
Image source: Shutterstock
