CoAL-RAG - Complexity-Aware Legal Retrieval-Augmented Generation Method - read online free

Sep 15, 2026

Uploaded by Hazel

Legal consultation questions show multi-level complexity, where one fixed retrieval strategy can over-reason for simple questions while providing poor interpretability for complex ones, harming answer quality and efficiency in high-stakes settings. This paper proposes CoAL-RAG, a complexity-aware legal retrieval-augmented generation method that builds adaptive routing via “question essence” and “retrieval consistency.” It quantifies reasoning demand from question logic structure, estimates complexity by the discrepancy between semantic retrieval and keyword retrieval, selects an optimal retrieval strategy, and filters context dynamically. Experiments validate strong improvements on Chinese and cross-jurisdictional generalization.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages15

Reading Time38 min

AccessFree to download

Introduction

Background and Challenges

Retrieval Strategies and Limitations

Proposed Approach and Contributions

Related Works

Legal Large Models

(Additional Sections)

arXiv : 2608 . 17536v2 [cs .CL] 24 Aug 2026 CoAL-RAG: A Complexity-Aware Legal Retrieval-Augmented Generation Method Jin Su 1 ,2 , Zhuofeng Zhao 1 ,2⋆, Huanhuan Wang 1 ,3 , and Hao Chen 1 ,3⋆⋆ 1 North China University of Technology, Beijing, 100144, China 2 Beijing Key Laboratory of Key Technologies for AI+ Domain Applications, Beijing, 100144, China 3 Beijing Key Laboratory on Integration and Analysis of Large-scale Stream Data, Beijing 100144, China [edzhao@ncut.edu.cn](edzhao@ncut.edu.cn) Abstract. Legal consultation questions exhibit multi-level complexity. A single retrieval strategy often leads to over-reasoning for simple questions and poor interpretability for complex ones, making it difficult to meet the requirements for both answer quality and efficiency in highrisk scenarios. To address this issue, this paper proposes CoAL-RAG, a complexity-aware legal retrieval-augmented generation method, which constructs a multi-dimensional evaluation mechanism based on “question essence” and “retrieval consistency” to enable adaptive routing of retrieval strategies. First, the reasoning demand is quantified according to the logical structure of the question. Then, the discrepancy between semantic retrieval and keyword retrieval is utilized to indirectly reflect problem complexity, thereby selecting the most appropriate retrieval strategy and dynamically filtering contextual information. Experimental results demonstrate that the proposed method significantly outperforms baseline models not only on Chinese legal benchmarks (SocialLawQA, LawBench) but also demonstrates strong cross-jurisdictional generalization on English datasets (LexGLUE, CaseHold) . Specifically, on Chinese datasets, the BLEU score improves by 42.5% and ROUGE-L reaches 3.6 times that of knowledge graph-based methods. On English benchmarks, CoAL-RAG maintains highly competitive accuracy, achieving an optimal balance between generation quality, deep logical reasoning, and system efficiency across different legal systems. Keywords: Legal Q&A · Retrieval-Augmented Generation · Complexity Awareness · Adaptive Retrieval · Knowledge Graph 1 Introduction Driven by breakthrough advancements of Large Language Models (LLMs) in Natural Language Processing (NLP) [4, 38], intelligent legal question answering ⋆ Corresponding author.⋆⋆ Equal contribution. 2 J. Su et al. (a) Evidence Gaps in Complex Reasoning Problems (b) Noise Introduction in Simple Factual Issues (c) Retrieval conflict Fig. 1. Limitations of Existing Methods is transitioning toward a new paradigm of semantic generation. Given the stringent demands for accuracy and traceability in high-stakes legal scenarios [35], integrating external knowledge bases can effectively mitigate model hallucination and knowledge obsolescence [16] . However, the complexity of legal consultation queries varies significantly. Questions such as “What is the statutory retirement age?” involve a single legal provision and require minimal reasoning, resulting in relatively low complexity. In contrast, queries such as “My employer fled without signing a labor contract after a workplace injury. Which laws have been violated and how can I protect my rights?” involve dense legal knowledge, strong conditional constraints, and multi-step reasoning, leading to substantially higher complexity [6] . Among existing retrieval strategies in the legal domain, pure vector-based retrieval lacks deep relational reasoning, yielding only fragmented legal provisions (Fig. 1a) . Conversely, the indiscriminate application of graph reasoning leads to computational redundancy and noise interference (Fig. 1b) . Furthermore, a direct hybrid of the two often results in inconsistent retrieval outcomesand reasoning conflicts (Fig. 1c) . To address these shortcomings, this paper proposes CoAL-RAG (Complexity-Aware Legal RAG) . Leveraging LangGraph, this method constructs a multi-dimensional evaluation mechanism centered on “quest

FAQs

What problem does CoAL-RAG address in legal retrieval-augmented generation?

CoAL-RAG addresses the mismatch between question complexity and retrieval strategy, which can cause over-reasoning on simple legal questions and weak interpretability on complex ones, reducing both quality and efficiency in high-stakes scenarios.

How does CoAL-RAG measure reasoning demand and complexity?

It quantifies reasoning demand using the logical structure of the question via a five-dimensional metric system, and estimates complexity indirectly by comparing semantic retrieval results with keyword (BM25) retrieval.

What are the reported performance benefits of CoAL-RAG across languages and jurisdictions?

On Chinese benchmarks (SocialLawQA, LawBench), BLEU improves by 42.5% and ROUGE-L reaches 3.6 times that of knowledge-graph-based methods. On English benchmarks (LexGLUE, CaseHold), CoAL-RAG remains highly competitive by balancing generation quality, deep logical reasoning, and system efficiency.

Sep 15, 2026

Uploaded by Hazel