LRAS: Advanced Legal Reasoning with Agentic Search

TL;DR

LRAS introduces a principled shift from static, closed-loop legal reasoning to dynamic Active Inquiry by coupling legal reasoning with agentic search. It identifies the introspection deficit and fragility of deep legal reasoning in existing LRMs, and presents two complementary learning modules: Introspective Imitation Learning and Difficulty-aware Reinforcement Learning. Empirical results show substantial gains across Chinese legal benchmarks, with average improvements from 8.2% to 32%, and strong robustness on out-of-distribution DiscLaw data. The framework enables autonomous, verifiable legal reasoning by grounding conclusions in external statutes and case law, advancing trustworthy AI for legal practice.

Abstract

While Large Reasoning Models (LRMs) have demonstrated exceptional logical capabilities in mathematical domains, their application to the legal field remains hindered by the strict requirements for procedural rigor and adherence to legal logic. Existing legal LLMs, which rely on "closed-loop reasoning" derived solely from internal parametric knowledge, frequently suffer from lack of self-awareness regarding their knowledge boundaries, leading to confident yet incorrect conclusions. To address this challenge, we present Legal Reasoning with Agentic Search (LRAS), the first framework designed to transition legal LLMs from static and parametric "closed-loop thinking" to dynamic and interactive "Active Inquiry". By integrating Introspective Imitation Learning and Difficulty-aware Reinforcement Learning, LRAS enables LRMs to identify knowledge boundaries and handle legal reasoning complexity. Empirical results demonstrate that LRAS outperforms state-of-the-art baselines by 8.2-32\%, with the most substantial gains observed in tasks requiring deep reasoning with reliable knowledge. We will release our data and models for further exploration soon.

Figures (7)

• Figure 1: Comparison of legal LLMs, traditional search, and our proposed legal agentic search framework. Conventional legal LLMs suffer from limited introspection capabilities, failing to recognize their knowledge boundaries. In contrast, LRAS systematically identifies knowledge gaps, retrieves relevant statutes with precision, and generates well-grounded legal conclusions.
• Figure 2: Impact of search strategies on Shallow and Deep Reasoning.
• Figure 3: The overall framework of our proposed method. The workflow consists of two main phases: Stage 1, Introspective Imitation Learning Data Curation, where raw legal data is processed, filtered for uncertainty, and synthesized into reasoning trajectories with search actions; and Stage 2, Difficulty-aware RL Data Curation, where hard samples are selected based on pass rates to train the model via reinforcement learning. The bottom panel illustrates the progressive training pipeline from the Base Model to the final LRAS RL Model.
• Figure 4: Analysis of search behaviors across model sizes. (a) Proportion of samples with $>1$ search actions. (b) Accuracy on samples requiring multi-step search. (c) Accuracy on samples with 0 searches.
• Figure 5: A case study comparing reasoning processes.