English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Harvey's Legal Agent Benchmark: Measuring AI on Real Legal Work

Forum topic · ✨步子哥 · 2026-08-09

Summary

Harvey AI has open-sourced the Legal Agent Benchmark (LAB), a new evaluation framework designed to measure how LLM-based agents perform on real legal tasks rather than on exam-style Q&A. The current task set is built around an M&A (mergers and acquisitions) data-room workflow, where an agent must ingest dozens of unstructured documents, follow an instruction set, use tools for retrieval and computation, and produce structured outputs. LAB uses an all-pass rubric scoring model: every rubric criterion must be satisfied for a task to count as complete, reflecting the legal context in which partial completion is unacceptable. An LLM-as-judge component evaluates outputs, and the release includes analysis of judge consistency and bias. The harness supports multiple model adapters (GPT-4, Claude, Gemini), tool use, batch sweeps, and detailed reports. The article highlights why legal work is harder to benchmark than coding, noting that rubric quality depends on domain expertise, and argues legal AI is shifting from feasibility questions to measurable quality comparisons.

What is LAB?

Harvey AI has open-sourced the Legal Agent Benchmark (LAB) to measure LLM agent performance on real legal work, not on legal Q&A. The current task set is an M&A (mergers and acquisitions) data-room scenario: the agent is given a data room full of documents, an instruction set, and access to tools (retrieval, calculation, file operations), and must produce structured outputs that are graded against an expert-authored rubric.

LAB has two parts:

1. Task dataset — agent instructions, related documents, and scoring rubrics. 2. Execution harness — runs the agent, evaluates results, and generates reports.

Repository: github.com/harveyai/harvey-labs Announcement: harvey.ai/blog/introducing-harveys-legal-agent-benchmark

Why Legal Work Is Hard to Benchmark

Legal work differs from coding work in three fundamental ways that make approaches like SWE-bench insufficient:

  • Inputs are documents, not code. Tasks involve dozens of pages of contracts, due-diligence reports, and regulatory filings. The agent must first understand unstructured text before acting.
  • Outputs are a quality spectrum, not pass/fail. A legal deliverable is graded on a continuous spectrum; LAB quantifies this with rubric items, but each item is still pass/fail and the overall score is a weighted aggregate.
  • Processes are non-linear. Agents must repeatedly search documents, cross-reference clauses, and revise earlier judgments, producing trajectories far more complex than a typical code task.
  • How LAB Scores Work

    LAB uses all-pass rubric scoring: a task has N rubric items and the agent must pass all of them to be marked complete. This is stricter than proportional grading because in a legal context a 90% pass rate may mean 10% of clauses went unreviewed, which is unacceptable.

    LAB also uses an LLM judge to evaluate outputs. The release includes explicit analysis of judge behavior across rubric items, including consistency and bias, because LLM judge quality directly affects benchmark validity.

    Harness Capabilities

    The execution harness supports:

  • Multi-model adapters — GPT-4, Claude, Gemini, and others can be plugged in.
  • Tool calling — document retrieval, file operations, computation.
  • Report generation — per-run reports showing each rubric item's pass/fail status.
  • Batch sweeps — run multiple tasks across multiple models for comparison.
  • This makes LAB an evaluation platform, not just a test set: it can be used for model comparison, capability tracking, and regression testing.

    Key Takeaways

  • Legal AI has shifted from "can it work" to "how well does it work." LAB provides a quantifiable answer to the quality question.
  • Agent capability matters more than raw model capability. LAB measures an agent's performance inside a legal workflow, where gaps in tool use, retrieval, and multi-step reasoning may exceed gaps in the underlying model.
  • Realistic environments matter. Tasks are built from real M&A data-room scenarios rather than artificial legal exam questions, so results should track real-world utility more closely.

An Open Question

LAB's rubrics are authored by Harvey's legal experts. That means benchmark quality depends on the designer's domain expertise. Unlike SWE-bench, where test cases come from real GitHub PRs and any developer can contribute, extending LAB to litigation or compliance requires domain experts. Harvey's design choice is to open-source the framework and task set while retaining expert review — a balance between openness and quality, and a structural feature of legal benchmarks more generally.

Tags

#legal-ai#ai-benchmark#llm-agents#harvey#m-and-a#open-source#evaluation#agentic-ai

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178630985