Researchers propose LEGO framework for legal reasoning
Researchers have introduced LEGO, a dual-module framework designed to improve how large language models handle complex legal reasoning. The system addresses two main limitations in current approaches: existing retrieval methods often miss the normative relationships between legal provisions, and standard Chain-of-Thought prompting may generate plausible but structurally unsound rationales.
Key points
- LEGO combines expert-annotated GraphRAG and structured Chain-of-Thought for legal reasoning.
- Using a Qwen3-8B backbone, LEGO achieves 40.53% exact-match accuracy on LawExamQACivil.
- The framework outperforms standard RAG and CoT baselines in the study.
LEGO combines Legal Expert GraphRAG, which uses an expert-annotated civil code graph to extract specific provision subgraphs, with Expert Chain-of-Thought, which structures reasoning into Provision-Fact-Conclusion steps. The framework utilizes a Qwen3-8B backbone model.
In testing, LEGO achieved 40.53% exact-match accuracy on the LawExamQACivil benchmark. This performance outperforms evaluated RAG and CoT baselines and is comparable to larger models, while maintaining robustness on multi-hop questions. The authors also report that LEGO yields the best results among evaluated baselines on open-ended benchmarks. Code and datasets are available on GitHub.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Anthropic's Claude discovers new enzyme system ART resembling crispr · 10 src
- MIT researchers publish book on visual AI for urban studies · 1 src
- Research shows evidence quality boosts fact‑verification scores · 1 src
- GPT OSS120B selects correct kinship terms 90.67% of the time in Hindi, Tamil, Korean · 1 src
- Experts use LLM disagreement to speed codebook revision · 1 src
Comments
via GitHub Discussions