Paper review · October 15, 2026
Legal-R1: Agentic Retrieval for Evidence-Grounded Legal Reasoning
Legal-R1: Agentic Retrieval for Evidence-Based Legal Reasoning
In a multi-agent setup with lawyer, paralegal and judge roles, the authors trained only the lawyer agent, which plans the evidence search, with SFT and GRPO, and an 8B model surpassed RAG baselines on Chinese legal benchmarks.
Problem
In legal question answering, a plausible answer is not enough; it must rest on valid, citable legal grounds. Standard RAG relies on a single semantic similarity search. As a result, it cannot select evidence by legal-specific criteria such as court authority, applicability and diversity of rulings.
Method
The system splits the work into three roles. At each turn, the Lawyer agent chooses one of three actions (statute search, case-law search or stopping) and writes or revises the search query. The Legal Assistant agent selects statutes with an LLM reranker that takes the search intent into account. Precedents are selected with MMR, which combines relevance and court-authority scores with a diversity penalty based on a penalty severity index (PSI), and are then summarized. The Judge agent receives the full evidence Trajectory and writes the final answer. Only the Lawyer is trained. It is first trained with SFT on Trajectories synthesized by a Teacher model, with the retrieved evidence tokens masked. GRPO is then applied with rewards for answer accuracy, Trajectory alignment, evidence usefulness, query quality and format.
Results
- Averages 60.28 across LawBench and LexEval, higher than same-size retrieval-based baselines R-Search (45.59) and A2search (42.28)
- Well above the backbone-class Qwen3-8B used without retrieval (27.85) and also above GPT-5.4-Mini (56.87), but below DeepSeek-V4-Flash-Preview (66.47)
- Scores 57.36 on LawBench's original metric and 56.54 on LexEval's original metric, higher than the legal-domain model LegalOne (23.25 / 35.30)
- In ablations, the authors report that post-training and authority- and diversity-based precedent reranking each contributed to the performance gain
Limitations
Evaluation is limited to Chinese legal benchmarks, and the precedent corpus contains only criminal judgments (LeCaRDv2). The 'LLM Judge' scores were graded by a single model, DeepSeek-V4-Flash-Preview, which is also one of the compared models. Because the reinforcement learning reward depends on alignment with Trajectories synthesized by the Teacher, the Teacher's biases may carry over directly.
Why it matters
In fields like law, where the authority and applicability of evidence matter, the work shows an approach that treats retrieval as a learned evidence-seeking policy rather than a fixed pipeline. Because retrieval planning is separated from answer generation, the trained Lawyer can also be paired with a different Judge model.
Key terms
- Agentic retrieval
- A retrieval method in which the model decides on its own, over multiple turns, whether to search, which sources to search, what queries to use and when to stop.
- MMR(Maximal Marginal Relevance)
- A reranking technique that picks documents one at a time, choosing ones that are highly relevant but do not overlap with documents already picked. Here, similarity in sentencing is used as the measure of overlap.
- PSI(Punishment Severity Index)
- A penalty severity index that converts different penalties, such as prison terms and fines, into a single number measured in days. It is used to measure how varied the outcomes are across precedents.
- GRPO
- A reinforcement learning method that samples multiple responses to one question and computes the advantage from each reward relative to the group average. It does not need a separate value network.
Original
Authors: Huaiwen Pang, Jiale Han, Junying Chen, Wenjun Xiao, Benyou Wang · Posted 2026-10-13 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.