Enoki: Efficient Multi-Level Hallucination Detection

On September 7, 2026, the research team proposed the Enoki framework, aimed at addressing the factual challenges of large language models in high-risk scenarios. This framework achieves statement-level verification and span-level localization by extracting text-anchored factual information as shared representations, avoiding the costly process of aligning propositions with segments. Enoki supports three information extraction modes: large language models, encoders, and rules, and balances accuracy with推理 costs through a unified interface. Experiments show that this framework remains competitive with less resource consumption and performs better in terms of fine-grained spans and entity-level localization. The research team also released the EnokiQA dataset, which includes annotations for alignment statement verification and span localization at both fine and coarse granularities.