Enoki Unifies Efficient, Fine-Grained Hallucination Detection

Large language models can produce convincing answers while quietly introducing unsupported facts. In high-stakes settings, a useful detector must do more than label an entire response as correct or incorrect. It should identify the factual unit that failed verification and show exactly where that unit appears in the generated text. The paper “Enoki: Efficient Multi-Level Hallucination Detection” presents a framework designed around this requirement.

Introduction: connecting verification with localization

Hallucination detection has commonly developed along two tracks. Claim-level systems decompose an answer into factual statements and determine whether each statement is supported by evidence. This produces interpretable verification units. Span-level systems instead mark suspicious words or phrases, which is more useful for editing, auditing, and user-facing feedback. The two views are complementary, but they are often implemented as separate pipelines.

Combining them can be expensive. LLM-heavy systems may need repeated calls for decomposition, verification, and alignment. Modular approaches can avoid some of that cost, but still require an additional mechanism to map a verified claim back to the precise text span that expresses it. Enoki addresses this gap with a shared intermediate representation.

Key points

• Relational facts act as the bridge. Enoki extracts open-information-extraction-style facts from the answer and anchors them to their source text. A fact can therefore serve as a unit for evidence checking while retaining its location in the response.
• Verification and localization share one path. The extracted facts are compared with supporting evidence. When a fact is unsupported, its text anchor allows the system to project the result back to the relevant span without a separate claim-to-span alignment stage.
• The interface is extractor-agnostic. Enoki can work with LLM-based, encoder-based, or rule-based extraction regimes. This lets users trade accuracy and inference cost without redesigning the rest of the detection pipeline.
• The target is finer-grained diagnosis. According to the paper’s reported results, Enoki remains competitive with strong claim-level systems while using fewer resources, and achieves stronger performance for span- and entity-level localization.
• A matching benchmark is released. EnokiQA contains annotations for both claim-level verification and span-level localization, enabling evaluation of the two capabilities in a coordinated setting.

Why it matters

The main contribution is architectural rather than merely classificatory. Enoki treats “is this fact supported?” and “which text should be flagged?” as two outputs derived from the same representation. That design is relevant to retrieval-augmented generation, automated document review, and high-stakes question answering, where a binary hallucination label is often insufficient.

The shared representation may also simplify deployment. A system can change the fact extractor according to its latency or budget constraints while preserving the verification and localization interface. This is particularly useful for teams that want to compare heavyweight LLM extraction with lighter encoder or rule-based alternatives.

The available material does not provide detailed benchmark scores, hardware measurements, or representative error analyses. The reported advantages should therefore be read as the authors’ overall experimental findings rather than as a complete cost-quality comparison. Further evaluation across domains will be important, since extraction quality and evidence availability can strongly influence downstream localization.

Even with that limitation, Enoki offers a clear direction for hallucination evaluation: represent generated text as evidence-linked facts, verify those facts, and retain enough grounding information to explain the result. This moves detection toward a unified workflow that is more interpretable and potentially more efficient.

Source: Hugging Face Daily Papers