Panoramic Study of Legal Large Language Models: In-depth Analysis of the LLM-and-Law Paper Repository

This article provides an in-depth analysis of LLM-and-Law, a continuously updated paper repository for legal large language models. Maintained by Jeryi-Sun, the repository is automatically updated daily (as of June 4, 2026). It systematically organizes research progress in seven key directions, including LLM applications in the legal field, legal reasoning models, and legal agents, offering valuable literature indexes for legal AI researchers. Its core value lies in integrating cutting-edge research, presenting a comprehensive view of the interdisciplinary field of legal LLMs, and helping to understand technical boundaries and possibilities.

Due to the professionalism, rigor, and reasoning complexity of its texts, the legal field has become the ultimate challenge for natural language processing. The rise of LLMs has brought new opportunities to legal AI, penetrating into areas such as judgment prediction, contract review, and legal consultation. The LLM-and-Law project emerged to systematically collect the latest achievements in this field and fill the gap in the integration of research literature on legal LLMs.

LLM-and-Law divides research into seven core directions:

1. Applications of large models in legal tasks (judgment prediction, document summarization, information retrieval, etc.);
2. Legal reasoning models (enhancing the legal reasoning capabilities of LLMs);
3. Legal agents (building intelligent systems that autonomously perform legal tasks);
4. Legal issues of large models (limitations such as hallucinations and copyright risks);
5. Data resources in the legal field (LexGLUE benchmark, LARGE evaluation tool, etc.);
6. Specialized legal large models (LawLLM, etc.);
7. Evaluation methods for legal LLMs (reasoning trajectory evaluation, etc.).

The core technologies in legal LLM research include:

1. Prompt engineering (legal syllogism prompts, precedent-enhanced prompts);
2. Retrieval-Augmented Generation (RAG, such as the KRAG framework and LARGE tool);
3. Knowledge graphs (aiding reasoning and hallucination detection);
4. Fine-tuning and continuous learning (domain pre-training, instruction fine-tuning, RLHF). These technologies aim to improve the legal adaptability and reliability of models.

Legal LLMs face four major challenges:

1. Hallucination issues (generating incorrect legal citations or interpretations);
2. Lack of interpretability (conflict between black-box characteristics and the auditability requirements of legal decisions);
3. Bias and fairness (amplification of biases in training data);
4. Cross-jurisdictional generalization (difficulty adapting to different legal systems). Response strategies include RAG anchoring to real documents, reasoning trajectory generation, bias auditing, etc.

LLMs are reshaping the landscape of the legal industry:

• For lawyers: Automating repetitive tasks, but low-end services may be replaced;
• For judges/arbitrators: Assisting in case retrieval, but cannot replace human judgment;
• For law students: A learning aid, but over-reliance should be avoided;
• For the general public: Lowering the threshold for legal consultation, but there is a risk of inaccurate advice.

The LLM-and-Law repository presents a comprehensive view of interdisciplinary research across computer science, law, linguistics, and other fields. Its daily automatic update mechanism ensures tracking of cutting-edge developments. Future directions should focus on making AI a legal assistant rather than a replacement, using technology to enhance the value of the rule of law, and exploring new paths to achieve justice.