acceptodds
Withdrawn from ICLR 2027

What Do Hallucination Benchmarks Measure? Exploratory Factor Analysis for Detection and Benchmark Diagnosis

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Hallucination detectors are usually evaluated in the same way, although they capture different properties of generated text. We present an unsupervised measurement framework that treats hallucination severity as a latent trait and uses a battery of indicators to distinguish input-conflicting (IC) hallucinations, which are unsupported by a source document, from context-conflicting (CC) hallucinations, inconsistent with the generated text itself. Within each IC and CC battery, heterogeneous indicators are rank-normalised and combined using a bifactor exploratory factor model, yielding a continuous score without hallucination labels, human-written reference outputs, or API calls. We evaluate seven benchmarks covering summarisation, source-grounded question answering, and document-level contradiction. On five IC-only benchmarks, source-grounding scores discriminate hallucination labels better than output-only coherence scores. Detection performance is competitive but varies markedly across datasets. Furthermore, Factor loadings show how strongly semantic similarity and source consistency contribute to the shared score. Across the benchmarks, detection is easier when faithful and hallucinated outputs differ in how much source wording they reuse, although our score provides information beyond this difference. After stratifying examples with similar levels of lexical novelty, our method retains more discriminative signal on benchmarks constructed from LLM-generated hallucinations than on extractive summarisation benchmarks, while generally adding predictive information beyond novelty alone. These findings suggest that current leaderboards can conflate factuality sensitivity with extractiveness. We therefore propose lexical-stratified evaluation and incremental-validity testing as complementary checks for hallucination benchmarks.

Primary area:
interpretability and explainable AI
Keywords:
Hallucination detection, Exploratory factor analysis, Hallucination severity estimation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_zkTbOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper reframes hallucination detection as a psychometric measurement problem: hallucination severity is treated as a latent trait measured imperfectly by a battery of reference-free indicators. Two indicator batteries are constructed—an input-conflicting (IC) battery (six source-grounded indicators) and a context-conflicting (CC) battery (eight output-only indicators). Within each battery, indicators are rank-normalised (RINT) and combined via an exploratory bifactor model with a Schmid–Leiman transformation, yielding a label-free continuous severity score. The framework is evaluated on seven benchmarks (AggreFact, FaithBench, HaluEval, RAGTruth, TofuEval, SummEval, ContraDoc). The central empirical claim is that detection performance is systematically confounded with lexical novelty: AUROC correlates strongly (ρ = .943) with the lexical separability of faithful vs. hallucinated outputs across benchmarks. The authors propose lexical-stratified evaluation and incremental-validity testing as standard complementary checks.

Critical strengths
  1. Valuable conceptual reframing. Treating hallucination detection as latent-trait measurement rather than prediction introduces a productive new lens—complete with reliability (ω_h), discriminant validity, and loading diagnostics—that the detection literature largely lacks.
  2. Strong statistical hygiene. Labels are used strictly post hoc for validation; factor count via parallel analysis; thresholds selected within training folds; Holm-corrected DeLong tests. This discipline is well above the norm for this area.
  3. A clean discriminant-validity negative control. The IC-vs-CC cross-battery test (Table 2, AUROC gaps of .052–.577) provides credible evidence that source-grounding and internal coherence are distinct constructs.
Critical weaknesses
  1. The central claim rests on n = 6. The headline finding that detector AUROC tracks lexical separability (ρ = .943, p = .005) is computed over six aggregate benchmark-level points. A single influential benchmark can drive it. This thin evidence is nevertheless stated as a general property of the field in the abstract and conclusions.
  2. Internal inconsistencies across tables. SummEval's configuration conflicts between Table 6 and Table 12; the HaluEval row in Table 14(b) appears truncated; Tables 2 and 9 are duplicative. These suggest insufficient proofreading of the quantitative record.
  3. Key failures are reported but not explained. The SummEval factor collapse and the ContraDoc near-chance performance each admit at least two competing explanations (construct mismatch vs. indicator insensitivity; prevalence/composition effects), and the paper does not disambiguate them—leaving the CC battery's validity unestablished on its only native benchmark.
Questions for authors
  1. Can the cross-benchmark novelty–AUROC analysis be replicated at the instance level (e.g., does within-dataset novelty stratification predict per-dataset score variation), rather than relying on six aggregate points?
  2. Please reconcile the SummEval configuration across Tables 6, 12, and 14, and supply the missing HaluEval entries in Table 14(b).
  3. How robust is "retained signal" to finer-grained novelty matching and to regression-adjusted (rather than matched) evaluation?
  4. For ContraDoc, can you provide evidence distinguishing "CC indicators measure a different construct than the annotation" from "CC indicators are insensitive"?
Reviewer_vXa5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes an unsupervised method for estimating hallucination severity by combining existing reference-free hallucination metrics. The paper treats them as indicators. The paper follows previous work and considers input-conflicting (IC) and context-conflicting (CC) hallucinations separately. To normalize the indicator scores, they apply a rank-based inverse normal transformation (RINT). They then use exploratory factor analysis (EFA), including Schmid–Leiman decomposition, to derive a general-factor severity score. The framework is fitted separately to each dataset and evaluated on seven hallucination benchmarks. Beyond detection performance, the authors investigate reliability, indicator loadings, construct validity, and the influence of lexical novelty on benchmark results. Detection performance varies considerably across benchmarks - strong results on AggreFact and competitive results on RAGTruth, but weaker performance on FaithBench, TofuEval, and ContraDoc. They also find that a simple bigram-novelty baseline performs surprisingly well on several datasets, and propose lexical-stratified evaluation and incremental-validity testing as additional checks for hallucination benchmarks.

Critical strengths
  • The paper combines existing hallucination metrics to derive a severity score without requiring labelled hallucination data.
  • The paper covers seven datasets spanning different tasks.
  • Experiments show that a simple bigram-novelty baseline achieves surprisingly strong performance, and the proposed lexical-stratified and incremental-validity checks are simple and useful for diagnosing benchmarks.
  • The paper performs extensive analyses and statistical tests to examine the results.
Critical weaknesses
  • The proposed framework essentially combines existing hallucination metrics into a single score using established statistical techniques. This limits the novelty of the work.

  • The factor model is fitted on a whole dataset, so the method cannot score a single new output without refitting. This limits its practical use as a detector.

  • The paper does not report the individual performance of the metrics on the same evaluation examples. It is unclear whether combining scores improves over the strongest individual metric for a particular dataset.

  • The general-factor score often performs similarly to or worse than a simple mean, and is worse on some datasets (Table 12).

  • The reported configuration differs across datasets (general factor, anchor factor, sentence-augmented score, or simple mean in Table 6), and the chosen one is usually the best-performing variant. No rule fixed before seeing labels is given for this choice, which weakens the label-free claim.

  • The factor models are fitted separately for each dataset, which limits evidence of generalization to unseen data.

  • The CC battery includes indicators such as repetition, fluency, and lexical diversity, which are not direct measures of self-contradiction.

  • The lexical analysis primarily controls for bigram novelty, while other factors such as output length, compression ratio, and source length may also affect detection performance.

Questions for authors
  • Given that the general-factor score often performs similarly to or worse than equal-weight averaging, what advantage does the proposed method provide?
  • Since the factor models are fitted separately for each dataset, can the learned factor generalize to unseen datasets without refitting?
  • Missing section reference in line 184
Reviewer_bbSNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper provides an unsupervised method to combine existing hallucination detectors into one score using factor analysis and shows that this unsupervised method is valid. Also, the paper provides a simple meta-checker, and find out that some hallucination benchmarks could be unreliable.

Critical strengths
  • The insights on the simple meta-checker is quite interesting that help identify the robustness on these hallucination benchmarks and how to improve them. This could guide future works on designing hallucination benchmark/checkers.
Critical weaknesses
  • Lack of any figures: Providing descriptive and simple figures and diagrams will be helpful to help readers understand the context.
  • The best result is the least convincing: The paper’s primary score (G) performs about the same as a simple average of the indicators (Table 12). The factor model is the paper’s main methodological contribution, yet it does not seem to add value. Also, the strongest result is on AggreFact, which the paper itself shows is one of the benchmarks where just measuring how much a summary copies the source already gives a high score. The main method adds little over a simple average.
Questions for authors

See weaknesses and the following:

  • Could you report the main results with one fixed configuration (e.g., G) for all datasets, so that the main table does not depend on picking the best configuration per dataset?
  • Could you check the ωh values in Table 7, especially those equal to 1.000? ωh should be below 1 when the indicators have any unique variance.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.