Self-evolving legal retrieval agent rewrites rules without model training

A new arXiv paper introduces a self-evolving agent for legal case retrieval. The system uses an LLM-based agent to create, test, and discard query-rewriting rules that improve BM25 retrieval without training model parameters.

The approach matters because legal retrieval still depends heavily on precise language and lexical alignment. Dense retrieval has improved, but BM25 remains a strong baseline in legal domains.

By letting rules evolve through validation experiments, the framework offers a lightweight way to improve search behavior while keeping the retrieval system interpretable.