Treffer: Multi-granular legal information fusion with adversarial compensation: A hierarchical and logic-aware framework for robust case retrieval.

Legal case retrieval plays a pivotal role in intelligent legal systems, requiring robust semantic understanding, legal logic modeling, and alignment with statutory reasoning. However, current approaches struggle to capture the hierarchical structure and causal dependencies inherent in legal documents. This issue is especially pronounced under sparse supervision or high information noise. To address these challenges, we propose HLAF-LCR, a Hierarchical and Logic-Aware Framework for Legal Case Retrieval, which systematically models the fact–reasoning–ruling structure of legal texts. We design a hierarchical dynamic masking mechanism that applies distinct masking strategies to different modules: detail-preserving masking in the fact module, high-ratio conceptual masking in the reasoning module, and causality-constrained masking in the ruling module. Furthermore, a cross-hierarchical semantic and logic-guided fusion mechanism enables contextual and logical information to propagate across modules, while a dual-channel adversarial compensation module leverages Wasserstein-based training to mitigate semantic distortion under aggressive masking. The framework is optimized using multi-task training. It jointly learns multi-granular representations, logic-guided reconstruction, retrieval relevance, and adversarial robustness. Extensive experiments on both Chinese and English legal case retrieval datasets — including LeCaRD, CAIL2022-LCR, COLIEE2020 and COLIEE2021 — demonstrate that HLAF-LCR significantly outperforms competitive baselines in both supervised and unsupervised settings. The interpretable model architecture and its robust performance under high masking ratios validate the framework's effectiveness in capturing fine-grained legal semantics, inference chains, and ruling logic. • A logic-aware framework models the fact–reasoning–ruling hierarchy in legal texts. • Dynamic masking preserves legal semantics across hierarchical document structure. • Cross-hierarchical fusion enables structured inference and logic consistency. • Adversarial compensation mitigates semantic distortion under high masking ratios. [ABSTRACT FROM AUTHOR]