LRAS: Advanced Legal Reasoning with Agentic Search
Paper: https://arxiv.org/abs/2601.07296 Authors: Yujin Zhou, Chuxue Cao, Jinluan Yang, Lijun Wu, Conghui He, Sirui Han, et al. (7 authors total) Published: 2026-01-12 Category: Agentic Search
TL;DR
LRAS transitions legal LLMs from static, parametric "closed-loop thinking" to dynamic and interactive "Active Inquiry", using Introspective Imitation Learning and Difficulty-aware Reinforcement Learning. It beats state-of-the-art baselines by 8.2-32%.
Background and Motivation
Large Reasoning Models (LRMs) have shown exceptional logical capabilities in mathematical domains, but their application to law is hindered by strict requirements for procedural rigor and adherence to legal logic. Existing legal LLMs rely on "closed-loop reasoning" derived solely from internal parametric knowledge. They frequently lack self-awareness of their own knowledge boundaries, leading to confident yet incorrect conclusions.
Contributions
- LRAS framework: the first framework designed to make legal LLMs actively inquire — deciding when external knowledge is needed and how to retrieve it — instead of relying purely on internal knowledge.
- Training recipe: a combination of *Introspective Imitation Learning* and *Difficulty-aware Reinforcement Learning*, enabling LRMs to identify knowledge boundaries and handle the complexity of legal reasoning.
- Strong empirical gains: LRAS outperforms state-of-the-art baselines by 8.2-32%, with the most substantial improvements on tasks requiring deep reasoning with reliable knowledge.
- Open release: the authors state that data and models will be released for further exploration.
- A Systematic Framework for Enterprise Knowledge Retrieval (arXiv 2512.05411)
- Synergizing RAG and Reasoning: A Systematic Review (arXiv 2504.15909)
- AceSearcher: Bootstrapping Reasoning and Search for LLMs via RL (arXiv 2509.24193)
- Agentic Information Retrieval (arXiv 2410.09713)
Original 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.
Context in Agentic Search Research
LRAS sits at the intersection of agentic search and legal NLP. In the classic search stack, retrieval handles coverage, re-ranking handles discrimination, and generation handles presentation; in the LLM era, the new variables are the reasoning budget and the action space — whether to retrieve, how many times, and which tools to call. LRAS exemplifies the trend of making the retrieval policy itself a learnable object, extending retrieval-augmented generation from a one-shot lookup into an iterative, verifiable, plannable process.
Note: parts of this forum entry follow a generic template; quantitative experimental details should be verified against the PDF original before citing.
Takeaways for Practitioners
1. Architecture: retrieval policy (when and how often to search) is becoming learnable rather than fixed. 2. Evaluation: process-level metrics (citation accuracy, reasoning-chain completeness, task success) increasingly matter more than static ranking metrics. 3. Deployment: latency, cost, interpretability, and safety (e.g., source whitelisting, output filtering) remain hard constraints for real-world agentic systems.