Enoki: Efficient Multi-Level Hallucination Detection
Enoki uses one anchored fact representation to verify claims and mark unsupported spans without a separate alignment step.
The paper frames that as a cheaper bridge between claim-level and span-level hallucination detection. Enoki can run with LLM-based, encoder-based, or rule-based extraction, trading accuracy against inference cost. The authors say it stays competitive with strong claim-level systems while doing better on fine-grained span and entity localization. They also release EnokiQA, a dataset with aligned claim verification and span localization labels. HF Daily Papers' note
The paper frames that as a cheaper bridge between claim-level and span-level hallucination detection. Enoki can run with LLM-based, encoder-based, or rule-based extraction, trading accuracy against inference cost. The authors say it stays competitive with strong claim-level systems while doing better on fine-grained span and entity localization. They also release EnokiQA, a dataset with aligned claim verification and span localization labels. HF Daily Papers' note
score 5