The rapid advancement of artificial intelligence has significantly accelerated the development of intelligent judicial systems, bringing profound transformations to both legal practice and legal research. As a major breakthrough in natural language processing, Large Language Models (LLMs) have demonstrated substantial potential in legal text analysis, legal reasoning, and intelligent decision support due to their powerful semantic understanding and generative capabilities. However, the intrinsic characteristics of the legal domain—such as its professional rigor, logical strictness, and normative authority—pose substantial challenges to the deployment of LLMs in real-world legal settings. In particular, issues related to interpretability, knowledge timeliness, reasoning consistency, and hallucination remain critical concerns. This survey provides a structured and systematic review of the applications and task-oriented research of LLMs in the legal domain. We first summarize the architectural characteristics, training paradigms, and technical development paths of representative legal-specific LLMs. We then examine three core judicial tasks—similar case retrieval, judicial examination question answering, and legal judgment prediction—by analyzing commonly used datasets, evaluation metrics, and methodological advancements. Furthermore, we synthesize the major challenges faced by legal LLMs across these tasks, including hallucination phenomena, insufficient reasoning consistency, and limited integration of domain-specific legal knowledge. Finally, we outline future research directions aimed at enhancing reliability, domain adaptability, and practical applicability, thereby providing guidance for subsequent research in legal artificial intelligence.
Peng et al. (Mon,) studied this question.