Self-Evolving AI Agent Boosts Legal Case Retrieval Accuracy

Legal case retrieval presents significant challenges due to the intricate nature of legal language and the necessity for precise lexical alignment between queries and relevant cases. While dense retrieval models have made progress, traditional methods like BM25 often remain strong baselines in this specialized domain. This observation motivated researchers to develop a novel self-evolving framework aimed at enhancing BM25 without requiring any parameter training. The proposed framework utilizes an LLM-based agent equipped with an automatic evaluation environment. This agent is designed to iteratively generate query rewriting rules, plan validation experiments to test various rule combinations, and then eliminate ineffective rules based on historical feedback. This continuous self-evolution process allows the system to refine its rule set dynamically. Evaluations conducted on the Chinese legal case retrieval benchmark LeCaRD-v2 demonstrated that this framework significantly outperforms non-evolutionary baselines, including human-designed rules and greedy rule selection. The performance gains were particularly notable when the framework was powered by a high-capacity core LLM. Detailed analyses revealed that the LLM's ability to leverage past experimental results and its intrinsic knowledge for rule elimination are crucial to the success of this self-evolving mechanism.