keepitsimple at SemEval-2025 Task 3: LLM-Uncertainty based Approach for Multilingual Hallucination Span Detection

Parameswari Krishnamurthy; Saketh Reddy Vemula

REVIEW 4 major objections 6 minor 8 references

keepitsimple at SemEval-2025 Task 3: LLM-Uncertainty based Approach for Multilingual Hallucination Span Detection

T0 review · 4 major / 6 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 Sampling an LLM 20 times exposes its hallucinated spans.

desk verdict Real mid-pack shared-task results, but the scoring function in §4 as written inverts the paper's own hypothesis—consistent spans get high scores, unmatched hallucinated spans get low ones. read the letter →

arxiv 2505.17485 v1 pith:3WG5YZXJ submitted 2025-05-23 cs.CL cs.AI

Saketh Reddy Vemula , Parameswari Krishnamurthy This is my paper · ORCID

classification cs.CLcs.AI

keywords hallucinationdetectionentropy-baseduncertaintystochasticsamplingmultilingualblack-boxLLMspanMu-SHROOM

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

The paper tries to establish that hallucinated spans in black-box LLM output can be located without any supervised training, by sampling the same prompt repeatedly and measuring how much the samples disagree. The working hypothesis is that a model that knows a fact will reproduce it consistently across samples, while a hallucinated fact will scatter into conflicting variants. The authors turn this into a span-level score by sliding a window over the generated text, matching each window against the sampled responses, and combining semantic entropy, lexical entropy, and a match-frequency penalty. On the 14-language Mu-SHROOM shared task the method ranks 18th overall, with a best single-language rank of 10th in Mandarin, and it requires no fine-tuning, making it cheap and applicable to any closed model.

What carries the argument

The load-bearing mechanism is a composite uncertainty score computed over the sampled response set. Semantic entropy $H_s$ treats the cosine similarities between a span's embedding and its matched spans as a probability distribution and takes its Shannon entropy; lexical entropy $H_l$ does the same over the frequency of matched spans; and the frequency score $F = 1 - |M_i|/|S|$ grows as matches become rarer. The final span score is $S_h(s_i) = 0.4H_s(s_i) + 0.4H_l(s_i) + 0.2F(s_i)$, and spans exceeding a per-language entropy threshold $\lambda$ are length-weighted-merged into the output. The informal logic is that a fact the model knows reproduces itself across samples (low entropy, many matches), while a hallucination fragments into inconsistent phrasings (high entropy, few matches).

What would settle it

A concrete check: take a set of passages where the model hallucinates a false 'fact' by confidently repeating the same wrong wording across all 20 samples (measured by near-identical outputs); if the detector does not flag those spans because their entropy is low, the consistency assumption is falsified. A complementary experiment would sweep the sampling temperature upward and show whether IoU and correlation improve or stay flat.

Watch

Extended reading notes

Core claim

The central discovery the paper argues for is that inconsistency across stochastic samples is a localizable signal: it marks not only whether a text is untruthful but where. Concretely, each candidate span of the generated text is matched to the most similar pieces among 20 low-temperature samples, and the spread of those matches is compressed into one score built from three terms: semantic entropy over embedding similarities, lexical Shannon entropy over matched spans, and a frequency term that penalizes spans with few matches. With weights 0.4, 0.4, 0.2 and a language-specific threshold, the score yields character-level hallucination spans that are merged when overlapping. The authors report consistent IoU and probability-correlation scores across languages, with the best numbers in Basque, Finnish, Italian, and Hindi, and note that the approach is deliberately training-free and black-box.

Load-bearing premise

The load-bearing premise is that sampling the same prompt 20 times at temperature 0.1 creates enough variety for hallucinated spans to look measurably more varied than faithful spans; the paper does not test how the detector behaves if the samples are nearly identical.

Editorial extensions

If this is right

A detector built this way transfers to a new language with no annotated examples: the same sampling step plus an embedding model and a tuned threshold are sufficient.
Because the score draws only on surface outputs, it works against closed black-box APIs where logits and hidden states are unavailable, a direct corollary of the method's design.
The per-language grid search over window size, stride, and threshold indicates that adapting the detector to a new language is a small hyperparameter exercise rather than a modeling exercise.
The high probability correlations reported in Italian and Hindi mean the score carries graded information: even when span boundaries are imperfect, the score tracks how strongly annotators believed a span was hallucinated.
The paper's best rank in Mandarin suggests the entropy signal survives typologically distant languages, which follows from the language-independent formulation of the score.

Reading between the lines

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

Implicit in the paper but left unexplored: the same entropy-over-samples construction could be applied to any future generative model without retraining, since the only model-specific piece is the sampling call itself.
The sampling temperature is fixed at 0.1 and never varied; the authors' own hypothesis predicts sharper entropy contrast at higher temperatures, which is a direct testable extension that the paper does not run.
The reported false positives stem from noise in the generated responses, so filtering or denoising the matched spans before entropy computation is a natural next experiment the paper only gestures at.
Because matching is lexical before it is semantic, hallucination spans that are paraphrased rather than lexically similar may be systematically underestimated; a test on paraphrase-heavy outputs would probe this boundary of the method.

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.

Desk Editor's Note

read the letter

Quick read on keepitsimple's Mu-SHROOM write-up. The reported results are real—mid-pack on an external shared task, with code and baselines—but the scoring function in §4, read literally, does the opposite of what the text claims. That's a load-bearing problem, not a quibble.

The genuinely new piece is the span-level reformulation: overlapping sliding windows over generated text, scored with a weighted mix of semantic softmax entropy, lexical entropy, and a match-frequency term, then thresholded and merged. That combination, applied to 14 languages, is not in the prior work they cite, and the multilingual evaluation is a fair contribution for a system paper. The paper is transparent about hyperparameters, includes error analysis, and ships a GitHub repository. Good.

The soft spot is formal. Eq. (5) computes Shannon entropy over a softmax of cosine similarities (Eq. (4)). If a span is consistent with all 20 sampled responses, every similarity is ~1, the softmax is uniform, and H_s = ln 20 ≈ 3.0. With α=0.4, β=0.4, γ=0.2, S_h ≈ 1.2, above every λ in Table 2 (0.5–0.7). A fully consistent span gets flagged as hallucinated. Conversely, if a hallucinated span has no match above threshold, M_i is empty, H_s=0, H_l=0, F=1, so S_h = 0.2, below every λ. The scoring function inverts the intended signal. The paper does not state that H_s is normalized by log|S|, that M_i is deduplicated, or that some other transform is applied before thresholding. Some undocumented step must be changing the computation, or the reported IoU/Cor numbers would not beat the mark-all baseline. The authors need to fix the equations or, if the code differs, say so.

The temperature 0.1 sampling (20 samples, top-p 0.9, top-k 50) is also a genuine concern: at that temperature responses are near-deterministic, so the entropy signal is weak. No sensitivity analysis is given. Missing details—embedding model, similarity threshold τ, deduplication behavior—make the results hard to reproduce from the paper alone. Those are secondary; the equation inversion is primary.

My verdict: the empirical claim is plausible because the shared-task evaluation is external, but the internal description is inconsistent. The reader's CONDITIONAL is fair; I'd lean toward requiring a correction before trusting the method description.

Recommendation: send it to review at a workshop venue, but with a direct instruction to reconcile §4 with the code. A serious referee would catch this quickly, which is exactly why it deserves referee time.

Referee Report

4 major / 6 minor

Summary. The paper describes a training-free, black-box system for the SemEval-2025 Mu-SHROOM multilingual hallucination-span detection task. The system generates 20 stochastic responses from Llama-3.2-3B-Instruct at temperature 0.1, slides a window over the original generated text, matches each window to spans in the sampled responses, and computes a weighted score combining semantic entropy, lexical entropy, and a frequency penalty; spans whose score exceeds a language-specific threshold are output as hallucinated. The authors report IoU and probability-correlation scores for 14 languages, a middle-of-the-pack average rank (18th), and a best rank of 10th for Chinese, and conclude that the uncertainty-based method accurately detects hallucinated spans without training.

Significance. If the mechanism worked as described, the paper would provide a simple, zero-training, language-agnostic baseline for hallucination span detection in black-box settings, building on SelfCheckGPT and semantic entropy. The paper's strengths are that it is evaluated on a public shared task with external gold labels, it reports per-language results against simple baselines, and it makes its code available. The reported performance is competitive but not state of the art. However, the formal statement of the scoring function in Section 4 is inconsistent with the stated hypothesis and with the reported thresholds, so the central claim cannot be verified from the manuscript as written.

major comments (4)

[§4, Eqs. (4)–(5), Table 2] The semantic entropy as written inverts the intended signal. For a fully consistent span si that is matched by all 20 sampled responses with sim(si,s'_j) ≈ 1, Eq. (4) assigns probability 1/20 to each response and Eq. (5) gives H_s = ln 20 ≈ 2.996. With α=0.4, β=0.4, γ=0.2, Eq. (8) gives S_h ≈ 1.2 even before H_l and F are added (and up to ≈2.4 if duplicates are counted as distinct symbols in H_l). This is above every λ in Table 2 (0.5–0.7), so the score would label the most consistent text as hallucinated, exactly contrary to the hypothesis in the abstract. The paper needs to state explicitly how H_s is normalized (e.g., by log|M_i|), whether M_i is deduplicated, or what additional transformation converts S_h into the probabilities used for the correlation metric.
[§4, Eqs. (7) and (10)] The frequency term cannot fire for the case it is designed to catch. If a hallucinated span appears in none of the sampled responses, M_i = ∅, so the sums in Eqs. (5)–(6) are empty (or zero by convention) and F(s_i) = 1; with γ=0.2, S_h = 0.2, which is below every λ in Table 2. Thus a span that is entirely absent from the samples is never predicted as hallucinated under the stated equations, even though F(s_i)=1 is supposed to indicate a high likelihood of hallucination. The authors should specify how missing matches contribute to the score (e.g., through a separate uncertainty term or a higher γ) and confirm that the submitted system actually implements this behavior.
[§4, §6] The paper never specifies the similarity threshold τ in Eq. (2), the sentence-embedding model E in Eq. (3), or the conversion from the span-level score S_h to the per-character predicted probabilities used for the Cor metric. Without these details, the reported IoU and Cor numbers cannot be reproduced or checked against the equations. At minimum, the τ value and the embedding model should be listed in Table 2 or in Section 5.2, and the probability conversion step should be described.
[§5.1] The central mechanism depends on stochastic sampling producing diverse responses for hallucinated spans, but the paper reports no sensitivity analysis for temperature, top-p/top-k, or the number of samples. At temperature 0.1 the model is nearly deterministic, and the formal issue in the first comment suggests that the entropy signal may not behave as intended; reporting the average diversity (e.g., pairwise similarity) of the 20 samples, and at least one experiment varying the sample count or temperature, would substantiate the claimed mechanism.

minor comments (6)

[§5.2, Table 2] The phrase 'extensive hyperparameter tuning' overstates the evidence; only 50 validation samples are available for each tuned language, and for Catalan, Czech, Basque, and Farsi the parameters are described as heuristic. This should be stated more cautiously.
[§4, Eq. (9)] The set O of 'overlapping spans' is not defined; specify the overlap criterion used for merging.
[§4, Table 2] The boundary-refinement step ('maximizing the entropy gradient at span boundaries') and the parameters MSL and BT from Table 2 are not defined in the equations; define them or remove them from the table.
[Table 4] The displayed strings contain missing spaces ('Di Stefanonel', 'strandenav'), and the red highlighting is not visible in the plain text; reformat the examples so the gold and predicted spans are clearly marked.
[§6] Calling correlation values in the range 0.16–0.39 'high correlation scores' is an overstatement; describe them as moderate or modest.
[Abstract and §1] The use of 'uniform' to describe consistent sampled responses is misleading, since in Eq. (4) a uniform distribution is the maximum-entropy case; rephrase to 'consistent' or 'identical'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the method's predictions are compared against external Mu-SHROOM annotations, and its assumptions are borrowed external hypotheses rather than restatements of the target labels.

full rationale

The paper's derivation chain is not circular. The hypothesis that consistent sampled responses indicate factual content and divergent responses indicate hallucination is explicitly taken from Manakul et al. (2023), an external prior work, and is tested rather than presupposed: the predicted hallucination spans are obtained by thresholding the score Sh(si) from Eq. (8)-(10), while the gold labels are human annotations from the Mu-SHROOM benchmark. No equation defines the gold labels in terms of Sh or vice versa, so the prediction is not equivalent to its input by construction. Hyperparameters such as w, t, and lambda are tuned on the validation split and then applied to the test split, which is standard model selection rather than fitting the test labels. The Limitations section explicitly concedes that supervised training data is under-used, which is a performance limitation, not a circularity. The formal scoring issue noted by the skeptic, namely that Eqs. (4)-(8) as written could assign high scores to perfectly consistent spans, is a correctness or soundness concern about the scoring function; it does not make the argument circular because the scores are still compared to independent external labels. There are no load-bearing self-citations, no imported uniqueness theorems, and no renaming of a known result as a new derivation. The paper's central claim therefore stands as an empirical, externally evaluated approach rather than a self-referential construction.

Assumptions & free parameters 7 free parameters · 3 assumptions · 0 invented entities

The method relies on a few domain assumptions rather than strong derivations: that response divergence signals hallucination, that lexical matching finds corresponding spans, and that entropy is a suitable divergence measure. All free parameters are either chosen by hand or tuned on small validation sets. No new entities are introduced.

free parameters (7)

Combination weights alpha, beta, gamma = alpha=0.4, beta=0.4, gamma=0.2
Chosen heuristically in Section 4, Eq. 8; not tuned on data.
Similarity threshold tau = not specified
Used to select matching spans in Eq. 2; no value is given anywhere in the paper, hampering reproduction.
Window size w = per-language: ar 4, de 4, en 5, es 4, fi 4, fr 4, hi 5, it 4, sv 4, zh 7
Tuned on validation splits for 10 languages; for Catalan, Czech, Farsi, Basque it is approximed heuristically (Section 5.2, Table 2).
Stride t = per-language: ar 2, de 2, en 3, es 2, fi 3, fr 2, hi 2, it 2, sv 2, zh 3
Tuned on validation splits for 10 languages (Table 2).
Entropy threshold lambda = per-language: ar 0.6, de 0.6, en 0.5, es 0.6, fi 0.6, fr 0.6, hi 0.6, it 0.7, sv 0.5, zh 0.6
Controls which spans are flagged; tuned per language on validation data (Table 2).
Minimum span length MSL = 3 for all listed languages
Post-processing parameter (Table 2).
Boundary threshold BT = 0.3 for all listed languages
Post-processing parameter (Table 2).

assumptions (3)

domain assumption Hallucinated spans cause stochastic responses to diverge more than faithful spans.
Stated in the Abstract and Section 1, this is the paper's foundational hypothesis, adopted from Manakul et al. (2023). It is assumed, not proven, and the entire entropy-based method depends on it.
domain assumption Lexical similarity matching with threshold tau reliably identifies corresponding spans in sampled responses.
Equation 2 assumes that for each window of the original text, matching text can be found in the sampled responses; if matching fails, entropy is computed over an empty or wrong set, invalidating the score.
standard math Shannon entropy and softmax over cosine similarities are valid measures of response divergence.
Equations 3 through 6 use standard mathematical definitions; their suitability for this detection task is asserted rather than derived from first principles.

how reviews work

0 comments

Cite this review

Pith. "Pith review of keepitsimple at SemEval-2025 Task 3: LLM-Uncertainty based Approach for Multilingual Hallucination Span Detection." pith.science (2026). https://pith.science/paper/3WG5YZXJ

@misc{pith2026250517485,
  author       = {Pith},
  title        = {Pith review of: keepitsimple at SemEval-2025 Task 3: LLM-Uncertainty based Approach for Multilingual Hallucination Span Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3WG5YZXJ}},
  note         = {Machine review of arXiv:2505.17485}
}

read the original abstract

Identification of hallucination spans in black-box language model generated text is essential for applications in the real world. A recent attempt at this direction is SemEval-2025 Task 3, Mu-SHROOM-a Multilingual Shared Task on Hallucinations and Related Observable Over-generation Errors. In this work, we present our solution to this problem, which capitalizes on the variability of stochastically-sampled responses in order to identify hallucinated spans. Our hypothesis is that if a language model is certain of a fact, its sampled responses will be uniform, while hallucinated facts will yield different and conflicting results. We measure this divergence through entropy-based analysis, allowing for accurate identification of hallucinated segments. Our method is not dependent on additional training and hence is cost-effective and adaptable. In addition, we conduct extensive hyperparameter tuning and perform error analysis, giving us crucial insights into model behavior.

Figures

Figures reproduced from arXiv: 2505.17485 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

8 extracted references · 5 linked inside Pith

[1]

Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. 2022. https://arxiv.org/abs/2108.07258 On the opportunities and risks of foundation models . Preprint, arXiv:2108.07258

arXiv 2022
[2]

Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. https://arxiv.org/abs/2407.21783 The llama 3 herd of models . Preprint, arXiv:2407.21783

arXiv 2024
[3]

Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. 2024. Detecting hallucinations in large language models using semantic entropy. Nature, 630(8017):625--630

2024
[4]

Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. 2025. https://doi.org/10.1145/3703155 A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions . ACM Transactions on Information Systems, 43(2):1–55

doi:10.1145/3703155 2025
[5]

Jannik Kossen, Jiatong Han, Muhammed Razzak, Lisa Schut, Shreshth Malik, and Yarin Gal. 2024. Semantic entropy probes: Robust and cheap hallucination detection in llms. arXiv preprint arXiv:2406.15927

arXiv 2024
[6]

Potsawee Manakul, Adian Liusie, and Mark J. F. Gales. 2023. https://arxiv.org/abs/2303.08896 Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models . Preprint, arXiv:2303.08896

arXiv 2023
[7]

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 Attieh, and Marianna Apidianaki. 2025. https://helsinki-nlp.github.io/shroom/ S...

2025
[8]

Yijun Xiao and William Yang Wang. 2021. https://arxiv.org/abs/2103.15025 On hallucination and predictive uncertainty in conditional language generation . Preprint, arXiv:2103.15025

arXiv 2021

Pith tools

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

[1] [1]

Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. 2022. https://arxiv.org/abs/2108.07258 On the opportunities and risks of foundation models . Preprint, arXiv:2108.07258

arXiv 2022

[2] [2]

Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. https://arxiv.org/abs/2407.21783 The llama 3 herd of models . Preprint, arXiv:2407.21783

arXiv 2024

[3] [3]

Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. 2024. Detecting hallucinations in large language models using semantic entropy. Nature, 630(8017):625--630

2024

[4] [4]

Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. 2025. https://doi.org/10.1145/3703155 A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions . ACM Transactions on Information Systems, 43(2):1–55

doi:10.1145/3703155 2025

[5] [5]

Jannik Kossen, Jiatong Han, Muhammed Razzak, Lisa Schut, Shreshth Malik, and Yarin Gal. 2024. Semantic entropy probes: Robust and cheap hallucination detection in llms. arXiv preprint arXiv:2406.15927

arXiv 2024

[6] [6]

Potsawee Manakul, Adian Liusie, and Mark J. F. Gales. 2023. https://arxiv.org/abs/2303.08896 Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models . Preprint, arXiv:2303.08896

arXiv 2023

[7] [7]

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 Attieh, and Marianna Apidianaki. 2025. https://helsinki-nlp.github.io/shroom/ S...

2025

[8] [8]

Yijun Xiao and William Yang Wang. 2021. https://arxiv.org/abs/2103.15025 On hallucination and predictive uncertainty in conditional language generation . Preprint, arXiv:2103.15025

arXiv 2021