LexChain: Modeling Legal Reasoning Chains for Chinese Tort Case Analysis

🌐 GitHub • 📚 AAAI • 🤗 HuggingFace

Legal reasoning is central to legal decision-making, yet current computational methods mostly rely on oversimplified frameworks (e.g., syllogism, IRAC), often tailored to criminal law. To address this, we introduce LexChain, a legally informed, structured reasoning framework that explicitly models the multi-step process of civil tort analysis in Chinese law.

We formalize the tort legal reasoning (TLR) task, build a high-quality benchmark (LexChain_eval), and propose a suite of reasoning-enhanced prompting and training strategies. Our findings show that structured legal reasoning significantly improves LLMs’ performance in tort case analysis and generalizes to broader legal AI tasks.

✨ The LexChain Framework

LexChain models tort case reasoning via a three-stage chain:

1. Legal Element Identification
   • Recognize plaintiffs, defendants, dispute types, and applicable statutes.
2. Liability Analysis
   • Reason over conduct, harm, causation, and fault.
   • Assign liability proportions and calculate damages.
3. Judgment Summarization
   • Produce legally grounded, structured decisions.

Each sub-task aligns with verifiable reasoning steps, enabling granular evaluation.

🧪 Benchmark: LexChain_eval

We curate 1,000 real-world tort cases from China Judgments Online.

• Covers 45+ dispute types (e.g., traffic, product, service liability)
• Annotated with: facts, parties, statutes, damages, liability, compensation, judgment and more
• Validated by legal experts

🖊️ Citation

If you use this work, please cite our AAAI 2026 paper:

@inproceedings{xie2026lexchain,
title={LexChain: Modeling Legal Reasoning Chains for Chinese Tort Case Analysis},
author={Xie, Huiyuan and Li, Chenyang and Zhu, Huining and Zhang, Chubin and Ye, Yuxiao and Liu, Zhenghao and Liu, Zhiyuan},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={40},
number={42},
pages={35913--35921},
year={2026}
}