When Rules Learn: A Self-Evolving Agent for Legal Case Retrieval

A new framework lets a large language model agent write and refine its own query-rewriting rules to improve BM25 legal case retrieval, without any parameter training, according to a paper submitted in 2026 [1]. Legal case retrieval is difficult because legal language is complex and demands precise lexical alignment between a query and relevant cases [1]. Dense retrieval models have made gains, but BM25 remains a strong baseline in the domain [1]. The proposed system equips an LLM-based agent with an automatic evaluation environment so it can iteratively create rewriting rules, plan validation experiments over rule combinations, and discard ineffective rules using historical feedback [1]. The method was tested on the Chinese legal case retrieval benchmark LeCaRD-v2 [1]. Results showed the self-evolving framework outperformed non-evolutionary baselines, including human-designed rules and greedy rule selection, especially when powered by a high-capacity core LLM [1]. The authors found that the LLM’s ability to leverage previous experimental results and its intrinsic knowledge of rule elimination were critical to refining the rule set through self-evolution [1]. Large language models are a class of generative AI system built on the transformer architecture, which has driven an AI boom since the 2020s [2][3]. Generative AI models learn patterns from training data and produce new text, images, or code in response to natural-language prompts [3]. The legal-retrieval framework taps this capability but constrains it with a rule-driven loop, avoiding the need for fine-tuning or retraining the underlying BM25 index [1]. Automated evaluation environments like the one used in the study are part of a broader trend in AI research toward systems that can self-improve through trial and error [2]. The approach echoes techniques seen in other domains, such as transfer learning across scientific datasets, where models trained on one corpus are adapted to boost performance on a smaller, related dataset [7]. In the legal context, the agent’s self-evolution replaces manual rule engineering with a feedback-driven process that the authors argue is more scalable [1]. The work was posted on arXiv, a preprint server that hosts papers across computer science and other fields, and was submitted in 2026 [1]. The paper includes detailed analyses of the mechanisms behind the self-evolution, showing that the LLM’s capacity to interpret past experimental outcomes is what drives the progressive refinement of rewriting rules [1].