GLIER: Generative Legal Inference for Case Retrieval

The newly introduced framework, GLIER (Generative Legal Inference and Evidence Ranking), aims to bridge the semantic divide between everyday user inquiries and formal legal documents in Legal Case Retrieval (LCR). Unlike traditional dense retrieval techniques that view LCR as a black-box semantic matching problem, GLIER approaches retrieval as an inference process involving underlying legal variables. This framework breaks the task into two stages focused on interpretability: first, a Joint Generative Inference module that converts raw queries into latent legal indicators (charges and legal elements) through a unified sequence-to-sequence approach; and second, a Multi-View Evidence Fusion mechanism that combines generative confidence with structural and lexical signals for accurate ranking. The study is available on arXiv under ID 2604.23779.