Pith. sign in

REVIEW 3 major objections 5 minor 15 references

Ask a Local: Detecting Hallucinations With Specialized Model Divergence

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Comparing language-specialized models' perplexities locates hallucinated spans in 14 languages without training or external data.

desk verdict A promising zero-training multilingual hallucination detector, but the central KL formula as written is undefined and blocks reproduction. read the letter →

arxiv 2506.03357 v1 pith:GV2IO5JO submitted 2025-06-03 cs.CL cs.AI

Aldan Creo , Héctor Cerezo-Costas , Pedro Alonso-Doval , Maximiliano Hormazábal-Lagos This is my paper
classification cs.CLcs.AI
keywords hallucinationdetectionmultilingualperplexitydivergencespecializedlanguagemodelszero-shotspantagginglargefactualcorrectness
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper introduces a hallucination detector that needs no training data and no external database, only a handful of language-specialized language models. The idea is that a model trained mainly on, say, Spanish should be more puzzled by a factual error about Spain than a model that knows nothing special about Spain; the gap in their per-word perplexities marks the error. The method averages each word's perplexity under specialized models into 'local' and 'foreign' numbers, takes a divergence between them, and turns high-scoring words into hallucinated spans with the help of an instruction-following model. On a human-annotated dataset in 14 languages, the detector scores Intersection-over-Union around 0.3, with Italian and Catalan reaching 0.42 and 0.38. The authors take this as evidence that the divergence signal transfers across languages without adaptation.

What carries the argument

The load-bearing object is the per-word hallucination score $H(w_i)=\beta\, \mathrm{KL}(\mathrm{PPL}_{\text{local}}(w_i)\,\|\,\mathrm{PPL}_{\text{foreign}}(w_i))+(1-\beta)\frac{1}{|M|}\sum_j \mathrm{PPL}(w_i;m_j)$, with $\mathrm{PPL}_{\text{local}}$ and $\mathrm{PPL}_{\text{foreign}}$ defined as weighted sums over specialized models. The weights $\alpha_j$ come from an instruction-following model asked to pick the most relevant specialization for the question-answer pair, with a temperature-smoothed softmax over the first token of each model's name. Perplexities are normalized to zero mean and unit variance using question tokens, then aggregated per word by taking the maximum of the word's token perplexities. A threshold $\sigma$ selects candidate words and the same instruction-following model proposes span boundaries; overlapping span probabilities are averaged. The specialized models are the Goldfish family for languages, with a general-knowledge model and a math model as optional extras, and an instruction-following model to assign weights and tag spans.

What would settle it

Re-running the 14-language evaluation with the divergence term replaced by a straightforward absolute difference $|\mathrm{PPL}_{\text{local}} - \mathrm{PPL}_{\text{foreign}}|$ — which is defined for scalars — would show whether the reported IoU of about 0.3 depends on the particular formula or on the domain-surprise intuition itself.

Watch

Extended reading notes

Core claim

The central discovery is that divergence between the perplexities of language-specialized models is a workable cross-lingual signal for hallucinated spans. Concretely, for each word $w_i$ the method builds $\mathrm{PPL}_{\text{local}}(w_i)$ as an $\alpha$-weighted average of perplexities from models specialized in the question's language/domain and $\mathrm{PPL}_{\text{foreign}}(w_i)$ as the complementary weighted average; the hallucination score is $\beta$ times the KL divergence between these two aggregates plus $(1-\beta)$ times the average perplexity. After normalizing perplexities using question tokens, words whose score exceeds $\sigma$ standard deviations above the answer mean are selected, and an instruction-following model converts them into complete spans. Tested on a human-annotated 14-language dataset, the method obtains IoU around 0.3 and comparable Spearman correlations, strongest on Italian (0.42) and Catalan (0.38), with no language-specific tuning.

Load-bearing premise

The load-bearing premise is that a specialized model is more surprised by errors in its domain, and that the paper's 'divergence' between two scalar perplexity values (written as a Kullback-Leibler divergence, though KL is not defined for scalars) detects that surprise.

Editorial extensions

If this is right

  • The detector needs no labeled hallucination data and no retrieval or external knowledge base, so it can be applied immediately to any language for which a specialized model exists.
  • Because performance stays near 0.3 IoU across all 14 languages without adaptation, the divergence signal appears language-agnostic, a property most trained detectors lack.
  • The best scores on Italian and Catalan suggest that the quality or number of specialized models per language directly affects detection, pointing to model coverage as a lever.
  • The span-tagging step produces complete meaning units (entities, clauses) rather than isolated words, matching the granularity human annotators use.
  • Using computationally light specialized models keeps the pipeline scalable to many languages; the extra general and math models are optional.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same divergence principle could transfer to non-linguistic specializations (code, medicine, law), where 'local' models would be domain experts rather than language experts.
  • The undefined KL on scalars suggests the numerical form in Equation 1 is likely a placeholder; a proper divergence defined over the distribution of perplexities across words might yield a more stable score.
  • The reliance on an instruction-following model for span proposals means the method is not fully unsupervised; an entirely automatic span-extraction heuristic would strengthen the zero-training claim.
  • The dependency on per-language specialized models implies the method's ceiling is bounded by model coverage and quality; for low-resource languages without such models, the signal weakens.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The manuscript proposes “Ask a Local”, a training-free hallucination detection method that compares perplexities of language-specialized models to identify hallucinated spans in multilingual LLM outputs. The method computes per-word perplexities with several specialized models, normalizes them using question tokens, and combines a weighted “local” versus “foreign” perplexity divergence with an average-perplexity term into a single score H(w_i). Hyperparameters are tuned on the English validation split, and the final method is evaluated on a 14-language human-annotated question-answer dataset, reporting IoU scores around 0.3 and Spearman correlations around 0.2–0.4. The authors argue that the approach generalizes across languages without external data or training and release code publicly.

Significance. If the method works as described, the paper offers a useful and reproducible contribution: it is training-free, does not require external knowledge sources, and explicitly targets multilingual hallucination detection, with code released and hyperparameters tuned on a separate validation set. The core intuition—that a domain-specialized model will be more surprised by errors in its domain—is attractive and worth testing. However, the central score equation is mathematically undefined as written, and the experiments do not isolate the divergence component from generic perplexity. These issues currently prevent the reported numbers from being attributed to the claimed mechanism. With a corrected formulation and a proper ablation, the contribution could be credible, but in its present state the central claim is not reproducible.

major comments (3)
  1. [Section 2.1, Eq. (1)] The hallucination score H(w_i) is defined as beta * KL(PPL_local(w_i) || PPL_foreign(w_i)) + (1-beta) * average perplexity, but PPL_local and PPL_foreign are scalars—weighted sums of scalar word-level perplexities in Eqs. (2) and (3). The Kullback-Leibler divergence is defined for probability distributions, not for two scalars. Therefore Eq. (1) is not computable as written. Since the tuned beta is 0.496, roughly half of the score is assigned to an undefined operation. The manuscript must either state the correct divergence formula (for example, a KL over a distribution of token or word perplexities, a log-ratio, or a squared difference) or point to released code that unambiguously defines the operation; otherwise the IoU and Spearman results in Table 3 cannot be tied to the stated divergence mechanism.
  2. [Section 3.1 and Section 4] No ablation isolates the divergence term. Setting beta=0 in Eq. (1) reduces the score to average perplexity alone, and setting beta=1 reduces it to the divergence alone. Reporting these two extremes, along with the tuned intermediate value, on the validation or test sets would show whether specialized-model divergence adds signal beyond generic perplexity. Without such an ablation, the paper's central claim—that divergence between specialized models drives hallucination detection—is not supported by the reported experiments.
  3. [Section 4] The baselines are too weak to support the performance claims. Comparing against a “neural baseline (close to 0)” and “marking all words as hallucinated” does not position the method relative to the state of the art, even though the manuscript itself cites Vázquez et al. (2025) as achieving scores around 0.6 on similar metrics. A comparison on the same dataset against existing / top-performing systems is necessary to substantiate the claim of “comparable Spearman correlation values” and the broader cross-lingual effectiveness of the method.
minor comments (5)
  1. [Section 2.2 heading] The heading “Addessing differences in tokenization” contains a typo; it should read “Addressing differences in tokenization.”
  2. [Section 2.1, Eq. (3)] Equation (3) has a stray “nX i” subscript after the summation symbol, which appears to be a leftover from LaTeX source; it should be removed or corrected.
  3. [Section 2.4] The text says the instruction-following model outputs probabilities over model-name tokens and that these probabilities are used to assign weights, but the subsequent softmax formula applies to logits divided by temperature. Please clarify whether the softmax is computed over logits or over token probabilities; these are different quantities.
  4. [Section 4, Table 3] The table reports a single scalar per language without confidence intervals or significance tests, despite only 200 examples per language; adding such information would strengthen the cross-lingual generalization claim.
  5. [Limitations] The Limitations section states that “our results are heavily influenced by the structure of these human annotations,” which is an important caveat but should also be reflected in the abstract and conclusion where the results are presented as consistent cross-lingual performance.

Circularity Check

0 steps flagged · score 0.0 of 10

No self-referential or fitted-input circularity; the method is self-contained and evaluated on a held-out test set.

full rationale

The paper's derivation is self-contained: the hallucination score H combines per-word perplexities from specialized models with weights assigned by an instruction-following model, and span tagging is a post-processing heuristic. No step defines its output in terms of its target: the score is not fitted to the human annotations, hyperparameters (beta, sigma, tau) are tuned only on the validation set, and results are reported on the held-out test set from Vazquez et al. (2025). There are no load-bearing self-citations; all cited prior work is external, and the dataset and baselines are independent. The undefined KL divergence between scalar perplexity values in Equation 1 is a genuine formal and reproducibility concern, but it is not circular: Equation 1 does not reduce to its inputs or to the annotated ground truth; it is simply not computable as written. The absence of an ablation isolating the divergence term is a weakness in evidence, not circular reasoning. Hence no circularity is present.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The method relies on a core behavioral assumption about specialized models, an undefined mathematical operation at the center of the score, and a weighting mechanism based on an instruction-following model. Hyperparameters are fitted on the validation set.

free parameters (3)
  • beta = 0.496
    Balance between divergence and average perplexity in Eq. (1), tuned on English validation set via Optuna.
  • sigma = 0.016
    Threshold for selecting words with high H scores, tuned on validation set.
  • tau = 3.393
    Temperature in softmax for model weights, tuned on validation set.
assumptions (3)
  • domain assumption Specialized models exhibit higher perplexity on domain-specific inaccuracies.
    Core hypothesis motivating the method, stated in Section 2.
  • ad hoc to paper The KL divergence between PPL_local and PPL_foreign as defined in Eq. (1) is a meaningful and computable quantity.
    As written, both are scalars, so standard KL is undefined; the paper implicitly assumes a computable form.
  • domain assumption The instruction-following model's next-token probabilities reflect the relevance of each specialized model to the QA pair.
    Weight assignment in Section 2.4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Ask a Local: Detecting Hallucinations With Specialized Model Divergence." pith.science (2026). https://pith.science/paper/GV2IO5JO

@misc{pith2026250603357,
  author       = {Pith},
  title        = {Pith review of: Ask a Local: Detecting Hallucinations With Specialized Model Divergence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GV2IO5JO}},
  note         = {Machine review of arXiv:2506.03357}
}
read the original abstract

Hallucinations in large language models (LLMs) - instances where models generate plausible but factually incorrect information - present a significant challenge for AI. We introduce "Ask a Local", a novel hallucination detection method exploiting the intuition that specialized models exhibit greater surprise when encountering domain-specific inaccuracies. Our approach computes divergence between perplexity distributions of language-specialized models to identify potentially hallucinated spans. Our method is particularly well-suited for a multilingual context, as it naturally scales to multiple languages without the need for adaptation, relying on external data sources, or performing training. Moreover, we select computationally efficient models, providing a scalable solution that can be applied to a wide range of languages and domains. Our results on a human-annotated question-answer dataset spanning 14 languages demonstrate consistent performance across languages, with Intersection-over-Union (IoU) scores around 0.3 and comparable Spearman correlation values. Our model shows particularly strong performance on Italian and Catalan, with IoU scores of 0.42 and 0.38, respectively, while maintaining cross-lingual effectiveness without language-specific adaptations. We release our code and architecture to facilitate further research in multilingual hallucination detection.

Figures

Figures reproduced from arXiv: 2506.03357 by the authors.

Figure 1
Figure 1. Our approach to detecting hallucinations. We propose a method to detect hallucinations in a model by comparing the perplexities of a set of specialized models. The perplexity of each specialized model is computed on the same input, and the divergence between the perplexities is used to detect hallucinations. are not directly comparable across models due to differences in tokenization. For instance, while one model’s… view at source ↗
Figure 2
Figure 2. Perplexity divergence. We show the normalized perplexity values of the words in the answer of [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

15 extracted references · 4 canonical work pages

  1. [1]

    AI@Meta. 2024. https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md Llama 3 model card

    2024

  2. [2]

    Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. 2019. https://doi.org/10.1145/3292500.3330701 Optuna: A next-generation hyperparameter optimization framework . In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD '19, page 2623–2631, New York, NY, USA. Association for Comp...

    arXiv 2019

  3. [3]

    Amos Azaria and Tom Mitchell. 2023. The internal state of an llm knows when it’s lying. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 967--976

    work page 2023

  4. [4]

    Chang, Catherine Arnett, Zhuowen Tu, and Benjamin K

    Tyler A. Chang, Catherine Arnett, Zhuowen Tu, and Benjamin K. Bergen. 2024. https://arxiv.org/abs/2408.10441 Goldfish: Monolingual language models for 350 languages . Preprint, arXiv:2408.10441

    arXiv 2024

  5. [5]

    DeepSeek-AI. 2025. https://arxiv.org/abs/2501.12948 Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning . Preprint, arXiv:2501.12948

    arXiv 2025

  6. [6]

    International Olympic Committee . 2025. P etra V A N S T A V E R E N . https://www.olympics.com/en/athletes/petra-van-staveren. [Accessed 07-02-2025]

    work page 2025

  7. [7]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K\" u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\" a schel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. In Proceedings of the 34th International Conference on Neural Information Processing Sy...

    2020

  8. [8]

    Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O ' Horo, Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, and Xian Li. 2022. https://doi.org/10.18653/v1/2022....

    doi:10.18653/v1/2022.emnlp-main.616 2022

Show all 15 references
  1. [9]

    Andrey Malinin and Mark Gales. 2020. Uncertainty estimation in autoregressive structured prediction. arXiv preprint arXiv:2002.07650

    2020 arXiv

  2. [10]

    Abhika Mishra, Akari Asai, Vidhisha Balachandran, Yizhong Wang, Graham Neubig, Yulia Tsvetkov, and Hannaneh Hajishirzi. 2024. Fine-grained hallucination detection and editing for language models. arXiv preprint arXiv:2401.06855

    2024 arXiv

  3. [11]

    Cheng Niu, Yuanhao Wu, Juno Zhu, Siliang Xu, Kashun Shum, Randy Zhong, Juntong Song, and Tong Zhang. 2023. Ragtruth: A hallucination corpus for developing trustworthy retrieval-augmented language models. arXiv preprint arXiv:2401.00396

    2023 arXiv

  4. [12]

    Gaurang Sriramanan, Siddhant Bharti, Vinu Sankar Sadasivan, Shoumik Saha, Priyatham Kattakinda, and Soheil Feizi. 2025. Llm-check: Investigating detection of hallucinations in large language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    2025

  5. [13]

    Ra\'ul V\'azquez, Timothee Mickus, Elaine Zosa, Teemu Vahtola, J\"org Tiedemann, Aman Sinha, Vincent Segonne, Fernando S\'anchez-Vega, Alessandro Raganato, Jindřich Libovický, Jussi Karlgren, Shaoxiong Ji, Jindřich Helcl, Liane Guillou, Ona de Gibert, Jaione Bengoetxea, Joseph...

    2025

  6. [14]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  7. [15]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.