Research map / Large language models / Reasoning and chain of thought

LexChain: Modeling Legal Reasoning Chains for Chinese Tort Case Analysis

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LexChain, a three-module legal reasoning chain for Chinese tort cases, builds a tort legal reasoning benchmark (LexChain_eval) and training data (LexChain_train), and shows prompting, SFT and DPO with LexChain-style reasoning improve LLM performance.

Problem

Existing computational legal reasoning relies on generic frameworks (syllogism, IRAC) and focuses on criminal cases, leaving civil tort reasoning under-modeled and under-evaluated.

Approach

A domain-specific three-module reasoning chain (legal element identification, liability analysis, judgment summarization) for Chinese tort law, operationalized into prompts, SFT data and DPO preference pairs, evaluated with a decomposed GPT-4o-judged rubric.

Key results

• DeepSeek-R1 with LexChain prompting reaches 60.84 overall on LexChain_eval, versus 45.36 zero-shot.
• Qwen-3-8B with SFT on LexChain data reaches 55.36 overall, versus 48.79 zero-shot, while syllogism SFT drops to 39.65.
• GPT-4o judge achieves 0.949 rating accuracy on 300 sampled instances; Cohen's Kappa 0.619 between annotators.

Surprising findings

• Syllogism-style SFT reduces performance below zero-shot baselines for all models tested.
• Labor services tort disputes are notably harder; GPT-4o scores below 35 in that category.

Limitations

• Benchmark built from Chinese tort judgments and limited to Chinese civil law context.
• Reasoning chains and gold data rely on DeepSeek-V3 generation with limited human validation.
• Inter-annotator agreement for judge validation is moderate (Kappa 0.619).

Experimental setup

Datasets: China Judgments tort court decisions (about 10,600 for training, 1,000 sampled for evaluation) Legal NER dataset Criminal Damages Calculation dataset

Benchmarks: LexChain_eval Legal NER Criminal Damages Calculation

Metrics: rubric score per dimension (0/1/2 or F1 for judgment) overall score

Baselines: GPT-4o o3-mini Claude-Sonnet-4 DeepSeek-V3 DeepSeek-R1 Qwen-3-8B InternLM-3-8B Llama-3.1-8B SFT_Syll

Builds on: syllogism-based legal reasoning IRAC DPO LoRA

Closest papers by meaning

• LexEval: A Comprehensive Chinese Legal Benchmark for Evaluating Large Language Models (NeurIPS 2024 · similarity 0.87)
• J&H: Evaluating the Robustness of Large Language Models Under Knowledge-Injection Attacks in Legal Domain (AAAI 2025 · similarity 0.84)
• Argumentative Large Language Models for Explainable and Contestable Claim Verification (AAAI 2025 · similarity 0.83)
• Lawma: The Power of Specialization for Legal Annotation (ICLR 2025 · similarity 0.83)
• Interpretable Long-Form Legal Question Answering with Retrieval-Augmented Large Language Models (AAAI 2024 · similarity 0.82)
• SaulLM-54B & SaulLM-141B: Scaling Up Domain Adaptation for the Legal Domain (NeurIPS 2024 · similarity 0.82)