Cross-Architecture LLM Ensembles, Feature-Based Reranking and Retrieval-Augmented Prompting for Legal Information Processing

This study reports on Team DU's participation in the COLIEE 2026 competition, covering five distinct legal information processing tasks: legal case retrieval, case entailment, statute retrieval, statute entailment, and legal judgment prediction. The researchers employed a diverse set of strategies tailored to the unique demands of each task. These included a learning-to-rank system for case retrieval, a three-stage pipeline with retrieval-augmented few-shot prompting for case entailment, and a cross-architecture ensemble of nine models for statute entailment.

The team achieved first place in Task 4 (statute entailment) using a cross-architecture ensemble of nine models, reaching 96.3% accuracy. In the Pilot Task (tort prediction), their unofficial multi-view system outperformed all official entries in verdict accuracy. For Task 2 (case entailment), the researchers discovered that simply adjusting the prompt from single-selection to multi-selection significantly improved F1 scores from 0.343 to 0.555, surpassing the competition winner. Similarly, in Task 3, replacing the entailment model with a larger Qwen3-235B model and using structured legal reasoning prompts boosted accuracy from 79.3% to 91.5% in post-competition analysis.

Legal documents present unique challenges, including extreme length, specialized terminology, and the need for precise reasoning over structural relationships. This paper highlights that there is no "one-size-fits-all" model for legal AI. Instead, success depends on matching the right inductive bias to the task—whether that is the structural and temporal features needed for case retrieval, or the multi-model consensus required for robust statute entailment. The findings suggest that prompt engineering and model selection are often as critical as the underlying architecture in high-stakes legal applications.