LRAS: Advanced Legal Reasoning with Agentic Search

Source: arXiv:2601.07296 — Yujin Zhou, Chuxue Cao, Jinluan Yang, Lijun Wu, Conghui He, Sirui Han, et al. (7 authors)

TL;DR

LRAS is the first framework designed to transition legal LLMs from static, parametric "closed-loop thinking" to dynamic and interactive "Active Inquiry", using agentic search to respect knowledge boundaries in legal reasoning.

Background and Motivation

Large Reasoning Models (LRMs) have demonstrated exceptional logical capabilities in mathematical domains, but their application to the legal field 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, and frequently lack self-awareness of their knowledge boundaries — leading to confident yet incorrect conclusions.

Approach

LRAS integrates two training techniques:

• Introspective Imitation Learning — teaches the model to recognize what it does and does not know.
• Difficulty-aware Reinforcement Learning — handles legal reasoning complexity via reward signals calibrated to task difficulty.

Together these enable LRMs to identify knowledge boundaries and invoke search as an interactive, agentic step rather than answering purely from memory.

Results

• LRAS outperforms state-of-the-art baselines by 8.2–32%.
• The most substantial gains appear in tasks requiring deep reasoning with reliable knowledge.
• The authors state they will release data and models for further exploration.

Original Abstract (verbatim)

> 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: Agentic Search

LRAS sits at the intersection of legal AI and agentic search, where retrieval becomes a learnable, sequential decision-making process (whether to search, how often, and which tools to call) instead of a one-shot pipeline step. It connects naturally with related work on RAG + reasoning synergies and reinforcement-learning-bootstrapped search agents.

Related Entries

• 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)

Takeaways

1. Knowing when not to trust parametric memory is as important as reasoning itself in high-stakes domains like law. 2. Combining imitation learning (for introspection) with RL (for difficulty calibration) is a practical recipe for agentic behavior. 3. Reported gains of 8.2–32% suggest knowledge-boundary awareness is a major bottleneck for legal LLMs.

*Note: quantitative details beyond the abstract should be verified against the full PDF.*