# Legal Research Bench ![Category](https://img.shields.io/badge/Category-Legal%20Agent-1f6feb) ![Metric](https://img.shields.io/badge/Metric-%25%20All--Pass-2da44e) ![Saturation](https://img.shields.io/badge/Saturation-Low-d29922) ![Created](https://img.shields.io/badge/Created-2026-555) [![Leaderboard](https://img.shields.io/badge/Leaderboard-vals.ai%2Fbenchmarks%2Flegal__research-0969da)](https://www.vals.ai/benchmarks/legal_research) [![arXiv](https://img.shields.io/badge/arXiv-2610.00609-b31b1b)](https://arxiv.org/abs/2610.00609) > **Quick answer:** Legal Research Bench is Vals AI's benchmark for agentic US legal research: an agent searches case law and the web, finds the governing statutes and precedent, and applies them to a fact pattern. An answer counts only if it meets every rubric item. Three models share the top score of 55.3% (Vals AI, October 2026). ## At a Glance **What it tests:** Multi-step legal research across federal and state law: finding statutes and case law, applying precedent to a set of facts, and synthesizing across jurisdictions or regulatory regimes. **Why it matters:** Legal research is where citation errors and missed precedent are costly. Law firms are listed as partners in evaluation, and the strict all-pass metric rewards complete answers rather than partly right ones. **Known limitations:** Vals AI lists the benchmark as proprietary and, like its other proprietary benchmarks, keeps the test questions private, so outside teams (Benchgen included) cannot reproduce the scores; they can only be mirrored. Grading is done by an LLM judge against a rubric, and the benchmark covers US law only. ## What Legal Research Bench Measures Questions span eight practice areas: administrative and regulatory, business and commercial, civil litigation, constitutional and civil rights, criminal, family, health, and immigration. The agent has case law search, web search and document retrieval, and has to research the question end to end before answering. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Agent / legal research | | Coverage | US federal and state law, 8 practice areas | | Agent tools | Case law search, web search, document retrieval | | Metric | % all-pass (every rubric item met); partial credit also reported | | Grading | LLM judge against an expert rubric | | Test set | Proprietary, questions not public | | Saturation | Low | | Created by | Vals AI | | Partners in evaluation | Fisher Phillips, McDermott, Reed Smith, LegalTech Hub | | Leaderboard | [vals.ai/benchmarks/legal_research](https://www.vals.ai/benchmarks/legal_research) | ## How Legal Research Bench Is Scored Each answer is checked against a rubric written for that question. Under the primary all-pass metric, a question counts as correct only if the answer satisfies every rubric item. A secondary partial-credit score gives the share of rubric points earned. An LLM judge does the grading. ## State-of-the-Art Results Vals AI leaderboard, all-pass, updated 2026-10-12: | Rank | Model | Score | Source | Date | |---|---|---|---|---| | 1 | Muse Spark 1.3 Max | 55.3% | [Vals AI](https://www.vals.ai/benchmarks/legal_research) | 2026-10 | | 1 | [Claude Opus 5](/models/anthropic/claude-opus-5) | 55.3% | [Vals AI](https://www.vals.ai/benchmarks/legal_research) | 2026-10 | | 1 | [Claude Fable 5.1](/models/anthropic/claude-fable-5-1) | 55.3% | [Vals AI](https://www.vals.ai/benchmarks/legal_research) | 2026-10 | | 4 | [Gemini 4 Argon](/models/google/gemini-4-argon) | 54.8% | [Vals AI](https://www.vals.ai/benchmarks/legal_research) | 2026-10 | | 5 | Wonder Full | 53.4% | [Vals AI](https://www.vals.ai/benchmarks/legal_research) | 2026-10 | Also on the Benchgen leaderboard: [Kimi K3](/models/moonshot-ai/kimi-k3) 44.2%, as reported by Moonshot AI in July 2026. ## Legal Research Bench on Benchgen > No Benchgen runs. The questions are not public, so the scores above are mirrored from Vals AI and model developers and cannot be re-run on Benchgen. ## Legal Research Bench vs Other Benchmarks | Benchmark | What it tests | Saturation | |---|---|---| | **Legal Research Bench** | Agentic US legal research with case law search | Low | | [Harvey Lab-AA](/benchmarks/artificial-analysis/harvey-lab-aa) | Legal knowledge-work quality | Low | | [DeepSearchQA](/benchmarks/google-deepmind/deepsearchqa) | Multi-hop web research & synthesis | Low | | [ResearchRubrics](/benchmarks/scale-ai/researchrubrics) | Rubric-graded research report quality | Low | ## Run Legal Research Bench on Your Model The official questions are private, so the leaderboard score cannot be reproduced outside Vals AI. On Benchgen you can build a rubric-graded legal research evaluation from your own questions and jurisdictions and track your agent against it over time. ## Frequently Asked Questions

What is Legal Research Bench?

Legal Research Bench is a Vals AI benchmark that tests AI agents on multi-step US legal research across eight practice areas, using case law search, web search and document retrieval.

What does a good score look like on Legal Research Bench?

The best all-pass score is 55.3%, shared by Muse Spark 1.3 Max, Claude Opus 5 and Claude Fable 5.1 (October 2026). Roughly half of the questions still defeat the best agents.

Who created Legal Research Bench?

Vals AI, with Fisher Phillips, McDermott, Reed Smith and LegalTech Hub as partners in evaluation. The paper is at arxiv.org/abs/2610.00609.

Can I reproduce the leaderboard scores?

No. Vals AI keeps the questions private, which keeps them out of training data but means only Vals AI can score them. Benchgen mirrors the published numbers.

Rank	Model	Score
1	gemini-4-carbon	55.77
2	claude-fable-5-1	55.3
3	claude-opus-5	55.3
4	claude-fable-5-1	55.3
5	claude-opus-5	55.3
6	gemini-4-argon	54.8
7	gemini-4-argon	54.8
8	kimi-k3	44.2

Legal Research Bench — Results

Legal Research Bench — Phases

Legal Research Bench