cs.CLOct 13, 2026
SaveLegal-R1: Agentic Retrieval for Evidence-Based Legal Reasoning
Organizations: The Chinese University of Hong Kong, Shenzhen · Shenzhen Loop Area Institute · Shenzhen Shenten Digital Technology Co., Ltd.
Abstract
Legal question answering requires answers that are not only plausible, but also grounded in legally valid and citable evidence. Standard retrieval-augmented generation often relies on one-shot semantic matching, which is insufficient when legal evidence must be selected according to authority, applicability, and diversity across heterogeneous sources. We introduce Legal-R1, an agentic retrieval framework for evidence-based legal reasoning. Legal-R1 decomposes legal reasoning into a multi-agent workflow, where a Lawyer agent learns when to retrieve, which legal source to query, what evidence to seek, and when the collected evidence is sufficient. A Legal Assistant agent retrieves from normative legal sources and precedent cases with authority-aware and diversity-aware selection, while a Judge agent generates the final answer from the complete evidence trajectory. We post-train only the Lawyer agent with masked supervised fine-tuning and trajectory-aware reinforcement learning. Experiments on multiple Chinese legal benchmarks show that Legal-R1 improves over general-purpose, legal-specific, and retrieval-augmented baselines.
Figures & tables
Figure 1: Motivation of Legal-R1. Legal-R1 replaces one-turn semantic matching with multi-turn evidence seeking over norms and cases.
Figure 2: Overview of the proposed agentic legal retrieval workflow.
Figure 3: Distribution of the training data across major legal task categories and fine-grained subtasks. Detailed statistics are provided in Appendix G .
| Model | Retrieval | Scale | Lawbench | LexEval | Avg | ||
|---|---|---|---|---|---|---|---|
| Trad. Score | LLM Judge | Trad. Score | LLM Judge | ||||
| LawGPT | ✗ | 7B | 11.33 | 13.76 | 9.49 | 9.18 | 10.94 |
| Luwen | ✗ | 7B | 30.06 | 32.77 | 14.03 | 16.97 | 23.45 |
| LegalOne | ✗ | 8B | 23.25 | 41.02 | 35.30 | 43.79 | 35.84 |
| Qwen3-8B | ✗ | 8B | 24.70 | 40.20 | 18.04 | 28.47 | 27.85 |
| A2search | ✓ | 7B | 46.82 | 49.14 | 35.23 | 37.94 | 42.28 |
Table 1: Model Performance Comparison on LexEval,and Lawbench Benchmarks. Judge by DeepSeek-V4-Flash-Preview.
Figure 4: Model Performance across Five Categories of LexEval.
Figure 5: Average Retrieval Turn for Different Complexity Tasks on Lawbench and Lexeval Benchmark.
| Model | Retrieval | Scale | Lawbench | LexEval | Avg | ||
|---|---|---|---|---|---|---|---|
| Trad. Score | LLM Judge | Trad. Score | LLM Judge | ||||
| Qwen3-8B | ✗ | 8B | 24.70 | 40.20 | 18.04 | 28.47 | 27.85 |
| Qwen3-8B w/ post-train | ✗ | 8B | 26.24 | 38.98 | 55.51 | 62.78 | 45.88 |
| Qwen3-8B w/ hybrid | ✓ | 8B | 52.02 | 59.85 | 50.52 | 61.49 | 55.91 |
| Qwen3-8B w/ rerank | ✓ | 8B | 53.28 | 57.97 | 50.12 | 57.31 | 54.67 |
| Legal-R1 w/ hybrid | ✓ | 8B | 57.08 | 63.69 | 55.47 | 61.96 | 59.55 |
Table 2: Ablation study: Model Performance Comparison on LexEval, Lawbench and average Benchmarks. Judge by DeepSeek-V4-Flash-Preview.
| Case Study: Autonomous Multi-Turn Retrieval | |
|---|---|
| Query | Will a debtor automatically be blacklisted if they fail to repay a debt within the court’s 25-day deadline? |
| Trajectory | Thinking Step 1: I need to locate the foundational statutory criteria for blacklisting. Retrieval 1: Article 1 . Thinking Step 2: Article 1 hinges on "fulfillment capability." I must retrieve its specific definition. Retrieval 2: Article 2 . Thinking Step 3: I need to search for procedural exemptions regarding the blacklisting process. Retrieval 3: Article 3 . Thinking Step 4: Let me invoke a precedent search to validate how courts apply these articles in practice. Retrieval 4: Precedent Case . |
| Answer | No, blacklisting is not automatic; according to Articles 1, 2, and 3, it strictly depends on whether the debtor willfully refuses to comply despite having the financial capability, provided they do not meet any legal exemptions. |
Table 3: A case study of Legal-R1’s autonomous multi-turn retrieval, with the full trajectory provided in Table 6 in the appendix.
| Legal-R1 Pipeline | LLM Judge | Trad. Score |
|---|---|---|
| Fixed 1-turn retrieval | 61.70 | 53.95 |
| Fixed 3-turn retrieval | 62.75 | 54.65 |
| W/ hybrid | 63.69 | 57.08 |
| W/o LLM rerank | 62.99 | 57.35 |
| W/o summarization | 61.44 | 52.95 |
| Full | 63.97 | 57.36 |
Table 4: LawBench ablations of retrieval rounds and evidence-processing components.
| System | nDCG | Recall | MRR | Hit Rate |
|---|---|---|---|---|
| text2vec | 0.3187 | 0.4109 | 0.2884 | 0.4109 |
| BM25 | 0.6993 | 0.7789 | 0.6725 | 0.7789 |
| Legal-R1 | 0.7659 | 0.8088 | 0.7516 | 0.8088 |
Table 5: Statute retrieval quality on 769 provision-oriented queries.
Figure 6: An idealized snapshot of Legal-R1’s autonomous thought-retrieval loop and final response synthesis.
| Model | Retrieval | Avg. Score |
|---|---|---|
| Qwen3-8B | ✓ | 54.67 |
| Qwen3-8B-SFT | ✓ | 59.98 |
| Legal-R1 | ✓ | 60.28 |
Table 6: Ablation study of training progression based on overall average scores.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: The prompt template utilized for normative evidence reranking. Within this template, {topk} denotes the predefined upper bound of retrieval turns, {query} and {context} represent the primary retrieval informational query, and {results} contains the candidate document pool obtained from the initial coarse-grained retrieval stage.
Figure 8: The prompt template utilized for precedent evidence summarization. Within this template, {fact_query} denotes the target case query to be investigated, while {doc_fact} , {doc_reason} , and {doc_result} represent the factual narrative, the judicial rationale, and the final adjudication outcome of the retrieved candidate precedent, respectively.
Figure 9: The prompt template utilized for judge answer generation. Within this template, {query} denotes the initial legal question inputted by the user, and {history} represents the comprehensive reasoning chain and multi-turn retrieval trajectory generated by the lawyer agent.
| Model | Avg. Score | LawBench Turns | LawBench Tokens |
|---|---|---|---|
| GPT-5.4-Mini | 56.87 | – | 399.23 |
| AutoRefine-Qwen2.5-7B-Instruct | 56.03 | 4.35 | 37226.08 |
| Qwen3-8B w/ hybrid | 55.91 | 0.35 | 5196.62 |
| Legal-R1 | 60.28 | 2.33 | 10861.85 |
Table 7: Accuracy–cost comparison with LawBench efficiency metrics.
Figure 10: An idealized snapshot of Legal-R1’s autonomous thought-retrieval loop and final response synthesis.
| Penalty Case | Calculation Process | PSI Result |
|---|---|---|
| 死刑 (Death penalty) | 100,000.0 | |
| 无期徒刑，剥夺政治权利终身 (Life imprisonment with lifelong deprivation of political rights) | 17,000.0 | |
| 有期徒刑十五年 (Fixed-term imprisonment of fifteen years) | 5,400.0 | |
| 有期徒刑三年，缓刑四年 (Fixed-term imprisonment of three years with a four-year suspension of sentence) | 216.0 | |
| 拘役六个月，并处罚金一万元 (Criminal detention of six months, with a concurrent fine of 10,000 RMB) | 224.0 | |
| 管制一年 (Public surveillance for one year) | 108.0 |
Table 8: Examples of PSI Calculation for Different Penalties.
| Main Category | Sub-category | Count |
| Knowledge Retrieval | Clause Recall | 500 |
| Clause Retrieval | 800 | |
| Subtotal | 1,300 | |
| Understanding | Case Similarity | 101 |
| Dispute Identification | 300 | |
| Concept Comprehension | 1,000 |
Table 9: Distribution of Dataset Categories and Sub-categories.