PaperlayerView exact revision on arXiv

Source-linked AI summary

Mycelium: A Meta-Study of Hallucination Detection Across Provenance, Typology, and Granularities

Julia Belikova, Elisei Rykov

arXiv:2610.33608v1cs.CLcs.AI

TL;DR

Hallucination detection scores are difficult to interpret across benchmarks because benchmark choices define different prediction problems. This paper organizes benchmarks by provenance, typology, and granularity and analyzes how those choices interact with detector evaluation. It concludes that reported performance is benchmark-dependent and that span metrics can respond quite differently to boundary variation.

  • Problem

    Benchmarks often define and measure hallucination differently, so scores from distinct prediction settings are difficult to compare directly.

    Source: §1 · p. 2

  • Method

    The paper analyzes hallucination detection as benchmark-conditioned measurement, using provenance, typology, and granularity to relate benchmark design to detector evaluation.

    Source: §1 · p. 2, §2 · pp. 2–3

  • Results

    Reported detector performance depends on benchmark conditions, while span metrics respond differently to boundary variation, including containment, overlap, and exact-boundary criteria.

    Source: §3.2 · pp. 6–7, §5 · p. 10

  • Takeaways & Limitations

    Interpret detector scores in light of benchmark conditions and make explicit which annotation variations should count as errors or remain invariant.

    Source: §5 · p. 10

    Span Coverage F1 can score 1.000 when predictions cover only 3.1% of reference characters, so containment alone does not show full localization.

    Source: §3.2 · p. 7

Abstract

from arXiv · show

Hallucination detection is increasingly important for evaluating and deploying large language models. However, existing methods are often compared across benchmarks that define and measure hallucination in substantially different ways, making reported scores difficult to interpret across settings. We study hallucination detection as a benchmark-conditioned measurement problem and organize existing benchmarks along three axes: provenance, which describes how model outputs and evidence conditions are obtained; typology, which captures the factuality and faithfulness failures represented by benchmark labels; and granularity, which specifies the unit at which hallucinations are annotated, predicted, and evaluated. Using this framework, we analyze how benchmark construction, annotation, and evaluation choices shape the underlying prediction task, and relate detector mechanisms and assumptions to the settings in which their performance is measured. We also examine span-level evaluation, where variation in annotation boundaries can substantially affect reported results, and propose Span Coverage F1 as a complementary containment-based metric. Our analysis clarifies what detector scores capture, highlights limitations of cross-benchmark comparisons, and motivates more explicit evaluation of generalization across benchmark conditions.

1 Introduction

Hallucination detectors are often evaluated on different prediction problems, so their headline scores are not directly comparable. The paper frames detection around benchmark conditions and analyzes how benchmark design, detector assumptions, and span-level evaluation shape what scores measure.

1 Introduction

  • Claim-, span-, and response-level evaluations measure different targets and depend on different decomposition, boundary, aggregation, and decision choices.Examples include FActScore and VeriScore, RAGTruth, and HaluEval or SHROOM, respectively.Source: §1 · pp. 1–2
  • Benchmark construction, labels, and evaluation determine the prediction target and constrain which generalization claims a reported score supports.Benchmarks differ in how outputs and evidence are obtained and whether annotations are binary or typed.Source: §1 · p. 2
  • The paper takes a benchmark-centered perspective, treating hallucination detection as a measurement problem structured by provenance, typology, and granularity.This perspective examines how benchmarks operationalize hallucination and what detectors are evaluated to recognize.Source: §1 · p. 2
  • The study analyzes benchmark construction and annotation, relates detector mechanisms and assumptions to benchmark conditions, and examines span-level metric choice under boundary variation.It proposes Span Coverage F1 as a complementary metric for contained boundary refinements.Source: §1 · p. 2

2 Scope and Formal Framework

The framework treats hallucination as a property of an evaluated configuration, not an intrinsic property of a model. It characterizes benchmarks by evidence and scope, three analytical axes, and the components that shape detector scores.

2 Scope and Formal Framework

  • A detector score characterizes the interaction between a method and its benchmark, rather than the method alone.The evaluated configuration includes the input, generated output, available evidence, and evaluation unit.Source: §2 · p. 2
  • The framework considers textual LLM outputs with declarative, truth-evaluable content and distinguishes source faithfulness from factuality.Source faithfulness concerns support from provided or retrieved context; factuality concerns consistency with time-indexed world knowledge.Source: §2 · p. 2
  • Granularity, provenance, and typology describe benchmark unit, output and evidence origin, and distinctions represented by labels.Granularity includes response-, span-, claim-, and multi-level formulations; provenance ranges from natural to elicited or synthetic; typology ranges from binary to typed labels.Source: §2 · p. 2, Figure 1 · p. 3
  • A benchmark comprises its instance distribution, provenance, granularity, label space, annotation rule, and evaluation metric.Detector outputs, evidence requirements, and model-access assumptions also condition how its score is interpreted.Source: §2 · p. 3
  • Scores across benchmarks are directly comparable only when target semantics, granularity, labels, alignment, metrics, and instance distributions are sufficiently compatible.Detector output may need alignment to benchmark representation through thresholding, span projection, claim decomposition, or aggregation.Source: §2 · p. 3

3 The Benchmark Landscape

The benchmark landscape treats hallucination detection as a measurement task shaped by provenance, typology, and granularity. Across these choices, annotations and evaluation rules define what detectors are being asked to recognize and how their scores should be interpreted.

3.1 Granularity

  • Granularity ranges from response-, span-, claim-, to multi-level targets; scores depend on annotation units, detector-to-benchmark alignment, and agreement metrics.Response-level judgments do not reveal error location or extent; span annotations vary in resolution and boundaries, claim scores depend on decomposition, and multi-level benchmarks require separate metrics.Source: §3.1 · pp. 3–4

3.2 Span Coverage F1: Evaluation under Boundary Variation

  • Span Coverage F1 combines prediction coverage and reference coverage as an instance-level containment measure, reflecting the assumption that a more specific subspan may be valid.It is intended for cases where an annotation marks an admissible region rather than a unique minimal boundary.Source: §3.2 · p. 6
  • Span Coverage F1 treats a narrower prediction contained within a reference as successful localization, but does not tolerate predictions extending beyond the annotated region.It is complementary to overlap-sensitive metrics; on partial overlaps its score falls to 0.347, and it is zero for disjoint annotations.Source: §3.2 · pp. 6–7
  • Span Coverage F1 rose from 0.410 with majority references to 0.750 with union references, while exact span F1 fell from 0.288 to 0.071.The same localization can receive different scores because metrics encode distinct evaluation questions, not a simple scale of strictness.Source: §3.2 · pp. 6–7

3.3 Provenance

  • Provenance distinguishes natural outputs, deliberately induced or selected semi-natural failures, and synthetic errors created through perturbation or generation.Semi-natural and synthetic construction can overrepresent induced conditions or construction artifacts, limiting assumptions about unbiased data or transfer to naturally occurring failures.Source: §3.3 · p. 7

3.4 Typology

  • Typology ranges from binary or coarse labels to evidence-relation and task-specific categories; recurring category correspondences are many-to-many, not exact equivalences.Their meanings depend on task, evidence regime, and annotation rules, so comparisons should account for category definitions and empirical distributions.Source: §3.4 · pp. 7–9

4 The Detector Landscape

The detector landscape is organized by generator access, detection signal, output granularity, and transfer through an inspector model. These choices interact with benchmark alignment, supervision, and calibration to shape what scores measure.

4 The Detector Landscape

  • The detector inventory distinguishes generator access, detection signal, output granularity, and transfer through a separate inspector model.These dimensions characterize the detector and, when needed, its alignment to a benchmark.Source: §4 · p. 9

4.1 Access and Signal Source

  • Black-box methods trade portability to closed models for possible external judge calls, retrieval, or repeated generation; white-box signals require open-model access and may be architecture-specific.Black-box methods vary in whether they use reference-free consistency signals or compare outputs with fixed or retrieved evidence.Source: §4.1 · p. 9

4.2 Output Granularity and Transfer

  • Localized predictions can usually be aggregated to sentence- or response-level scores, but response-level decisions cannot become faithful spans or claims without localization or decomposition.For claim-level detectors, decomposition choices affect verified units and the score denominator, so results may conflate extraction and verification quality.Source: §4.2 · p. 9

4.3 Supervision and Calibration

  • Prompted judges and sampling methods may need no training, while probes and fine-tuned detectors require labeled supervision; greater supervision often improves in-distribution accuracy but reduces portability.Comparisons should report supervision source, calibration protocol, and aggregation rule with the detector score.Source: §4.3 · pp. 9–10

5 Conclusion

The study concludes that detector performance is inseparable from benchmark construction and evaluation choices, which determine what the score represents. Its span-metric analysis illustrates this dependence, motivating explicit testing of robustness across benchmark conditions.

5 Conclusion

  • Benchmark choices in generation, categorization, localization, and scoring determine what detectors are evaluated to recognize and what reported performance represents.The same localization can score differently under exact-boundary, overlap, or containment criteria, and nested boundary disagreement is common in multi-annotator Mu-SHROOM data.Source: §5 · p. 10
  • Span Coverage F1 accepts contained refinements, but its stress-test failure shows that no single metric captures every aspect of localization quality.The study calls for explicit choices about which annotation variations should be invariant and controlled evaluation across construction, error type, annotation unit, and metric semantics.Source: §5 · p. 10

Related papers

Loading 2610.33608v1…