Team DU participated in all five tasks of the COLIEE 2026 competition, utilizing open-weight systems for legal case retrieval, entailment, statute retrieval, and judgment prediction. The team achieved first place among 33 submissions in Task 4 (statute entailment) with a cross-architecture ensemble of nine models reaching 96.3% accuracy.
- Task 4: A cross-architecture ensemble of nine models from three families achieved 96.3% accuracy, placing first among 33 submissions.
- Pilot Task: A multi-view system combining five claim-level models and refining verdicts with claim-derived features scored 73.1% TP accuracy and 68.2% RE F1.
- Task 2: Switching prompts from single- to multi-selection raised F1 from 0.343 to 0.555, exceeding the best official submission.
- Task 3: Replacing the entailment model with Qwen3-235B and using a structured legal reasoning prompt increased accuracy from 79.3% to 91.5%.
- Task 1: A learning-to-rank system combining lexical and semantic retrieval with 34 structural features achieved F1 = 0.314.
The results demonstrate that legal information processing benefits from different inductive biases across tasks, with cross-architecture ensembling, feature-based reranking, and retrieval-augmented prompting proving effective in specific settings.