Retrieval-Augmented Legal Judgment Prediction with Adaptive Case Retrieval

Legal judgment prediction (LJP) models are mostly parametric, although judges routinely consult precedents, and recent retrieval-augmented LJP systems attach a fixed number of precedents to every case. We propose adaptive case retrieval (ACR), which treats a calibrated parametric predictor and the precedent votes of a sparse and a dense retriever at seven retrieval depths as experts, and learns a gate that mixes them from confidence, similarity, label-dispersion and agreement features. The dense retriever is a judgment-consistent contrastive (JCC) encoder whose objective adapts the contrastive knowledge-infusion loss of LegalReasoner by treating cases with agreeing judgments as positives. On CAIL2018-Small with train–test duplicates removed (112 charges, 91 law articles, 11 term intervals), JCC raises 10-nearest-neighbour charge accuracy from 76.9% to 90.7%, and ACR improves a TF-IDF+SVM predictor by 0.62 and 1.28 accuracy points on law-article and prison-term prediction (p < 0.001), matching the best validation-tuned fixed-k interpolation in distribution. Its advantage grows when the precedent memory expands without re-training: with every component fitted on 20% of the labels, expanding the memory to the full training set raises term accuracy by 1.87 points under ACR versus 0.53 points under the fixed scheme, recovering 86.8% of the gap to a fully re-trained model, and the gains do not stem from near-duplicate precedents. A confidence-based routing rule bypasses retrieval for the 70–90% most confident cases while preserving the full charge and article gains, and the complete pipeline costs about 11 ms per case on one CPU core.