NOWJ's adaptive pipelines win legal AI benchmarks at COLIEE 2026

Researchers from the NOWJ team (affiliated with Japanese institutions) have released a comprehensive paper detailing their methodologies across all five tasks of the COLIEE 2026 competition—a premier benchmark for legal information retrieval and reasoning. For Task 1 (Legal Case Retrieval), they built a four-stage pipeline: candidate filtering with BM25, dense retrieval using two complementary embedding models, cross-encoder reranking via fine-tuned generative rerankers and MLP-based pairwise classification, and an adaptive per-query cutoff predictor that dynamically decides how many results to keep. This adaptive cutoff is a novel contribution that significantly improves precision-recall balance.

For other tasks, the team employed a mix of traditional and modern techniques. Task 2 (Legal Case Entailment) used BM25 filtering, T5-based reranking, and LLM-based entailment verification with consensus ensemble. Task 3 (Statute Law Retrieval and Entailment) adopted a retrieval-augmented generation (RAG) framework with dense retrieval, attention-based reranking, and few-shot-prompted LLM reasoning. Task 4 (Legal Textual Entailment) introduced a dynamic routing pipeline that classifies query difficulty and dispatches to either a balanced few-shot solver or a structured zero-shot chain-of-thought solver. The Pilot Task (Legal Judgment Prediction) combined hierarchical transformers with CRF layers, argument relation mining, and probabilistic argumentation graph reasoning. The paper is presented at COLIEE 2026 and available on arXiv (2607.16603).