Human-Centered Hybrid Intelligence: A Specialized Generalist Model for Intellectual Property Case Retrieval in Judicial Decision Support

In response to the challenges of result precision, output traceability, and interaction efficiency in intellectual property (IP) case retrieval for judicial decision-making, this study develops a hybrid retrieval framework. Rather than pursuing algorithmic performance alone, the framework incorporates legal-domain structure into model design. Accordingly, a Specialized Generalist Model (SGM) for IP case retrieval is designed. Methodologically: (1) Based on common elements in IP adjudication, a “Right Basis – Infringing Act – Damages Basis” triadic feature extraction mechanism is constructed to support structured semantic parsing of legal texts; (2) A dynamic threshold-based gated scheduling module coordinates a domain-specialized model and a general reasoning model through a dual-threshold verification strategy; (3) To address the long-text characteristics of IP cases, a paragraph–sentence dual-layer encoding domain-specialized model is built; (4) A chain-of-thought guiding mechanism is used to organize structured feature extraction and result verification in the general reasoning model. Experimentally, we construct IP-CaRD v1.0, a benchmark dataset for IP case retrieval covering 240 query cases across 53 fourth-level causes of action with expert annotations. Results show that SGM achieves an F1@3 score of 0.9149, outperforming the best baseline SAILER by 11.6%, and the domain-specialized model reaches a positive/negative discrimination rate of 90.41%. Human evaluation suggests that SGM receives higher perceived result usability, interaction usability, and explanation quality scores than the compared baselines. These findings suggest that SGM provides a design reference for human-centered interaction and human-AI collaboration in specialized legal retrieval, while broader workflow validation remains necessary before deployment.