Hallucination Detection with Small Language Models

Ming Cheung

REVIEW 4 major objections 7 minor 38 references

Hallucination Detection with Small Language Models

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Sentence-level yes-probability scores from two small language models detect hallucinations in retrieval-augmented answers better than a ChatGPT-based judge.

desk verdict A sensible small-model ensemble idea, but the claimed 10% F1 gain is not supported because the evaluation tunes the threshold and normalization on the same data. read the letter →

arxiv 2506.22486 v1 pith:54PHBTB5 submitted 2025-06-24 cs.CL cs.AI

Ming Cheung This is my paper

classification cs.CLcs.AI

keywords hallucinationdetectionsmalllanguagemodelsretrieval-augmentedgenerationanswerverificationsentencesplittingP(yes)probabilitymodelensembleF1score

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 argues that hallucinated answers in retrieval-augmented generation can be caught by small language models rather than by large, closed, or API-only ones. The proposed framework splits a generated answer into sentences, asks two small models whether each sentence is supported by the retrieved context, and turns the probability of a 'yes' first token into a single score via normalization, averaging, and the harmonic mean. On a dataset of HR-policy questions with correct, partially correct, and wrong answers, the framework reports F1 scores about 10 percent higher than the ChatGPT-based baseline and 6.6 percent higher than prompting a single small model on the whole response. The payoff would be that cheap, locally hosted models can serve as a scalable verifier for LLM outputs.

What carries the argument

The machinery is a three-stage scoring pipeline. A splitter segments the answer into sentences $r_{i,j}$; each small model $m$ assigns a score $s^{(m)}_{i,j} = P(\text{token}_1 = \text{yes} \mid q_i, c_i, r_{i,j})$, the probability that the first generated token is 'yes' when asked to verify the sentence against the context. Scores are normalized per model as $(s^{(m)}_{i,j} - \mu_m)/\sigma_m$ to put different models on a common scale, averaged across the $M$ models, and finally combined over sentences by the harmonic mean (restricted to positive values) to yield the response-level score $s_i$. This score is the object whose distribution separates correct, partial, and wrong responses.

What would settle it

A reader could split the dataset by question, estimate $\mu_m$, $\sigma_m$, and the threshold on the first half, and score the second half; if the held-out F1 for detecting correct from partial responses does not stay near 0.81, the core claim of a 10 percent improvement is not supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that $P(\text{token}_1 = \text{yes} \mid \text{question}, \text{context}, \text{sentence})$, estimated locally from multiple small models and combined across sentences, is a more reliable hallucination signal than asking one large model to judge the whole answer. For each sentence, the framework normalizes each model's yes-probability by its own mean and standard deviation, averages the normalized scores across models, and takes the harmonic mean over sentences to produce a final score. Thresholding this score separates correct answers from wrong ones almost perfectly and from partially correct ones with the best reported F1 of 0.81, outperforming ChatGPT and single-model P(yes) baselines. The authors take this as evidence that multiple small models can effectively verify answers when ground truth is absent.

Load-bearing premise

The framework assumes that the per-model normalization moments and the decision threshold are available without being fit to the same responses that are later scored; if they are tuned on the evaluation data, the reported F1 values are optimized, not predictive.

Editorial extensions

If this is right

If the framework holds, organizations can run hallucination detection entirely on local small models, avoiding API costs and data-exfiltration concerns.
Sentence splitting is central: the same averaging process fails to separate partial from correct answers when applied to whole responses, so per-sentence checking is what buys the improvement.
Combining multiple small models outperforms either model alone, implying that model diversity, not scale, contributes to verification accuracy.
The harmonic mean over sentences outperforms geometric, arithmetic, min, and max aggregators on partial answers, suggesting that the weakest-sentence penalty is the useful inductive bias.
The approach transfers to any domain where question, answer, and retrieved context are available, not just HR-policy questions.

Reading between the lines

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

Beyond the paper: if the normalization moments and decision threshold are estimated on held-out data rather than on the evaluation set, the reported F1 gains are likely to shrink, because the paper does not describe a train/test split and the text only says the best thresholds were the ones yielding the highest F1.
Beyond the paper: the same normalized-yes-probability pipeline could be applied to closed-book fact-checking of summaries, replacing a large judge model with a cheap ensemble.
Beyond the paper: a testable extension is to replace the fixed harmonic-mean aggregator with a learned weighting of sentence scores, which could improve detection of partially correct answers where the current best F1 is 0.81.
Beyond the paper: measuring the calibration of $P(\text{token}_1 = \text{yes})$ for each model would reveal whether the ensemble gain comes from better score separation or from averaging away per-model miscalibration.

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

Short version: the proposed method—two small models, sentence splitting, per-model normalization, harmonic mean—is a sensible, cheap-to-run idea, and the paper explains it clearly. But the one number that matters, the 10% F1 improvement, is not trustworthy as reported. The threshold is chosen on the evaluation set (Section V-D says 'thresholds yielding the highest F1 scores'), and Eq. 4 says the normalization moments 'can be computed based on previous responses' without saying what was done. So we don't know if the F1s are in-sample optima. With only about 100 question/context/answer triples, that's a real problem.

What's actually new: the specific combination of multiple small models with normalization and harmonic mean over sentence scores. Each ingredient is known—P(True), semantic entropy's sentence splitting, small models—but the ensemble-of-verifiers framing is a modest extension, not a new capability. The paper is honest about its scope: it's a practical verification tool for RAG, not a theory of hallucination. The writing is straightforward and the equations are simple and internally consistent.

Soft spots, in order of severity. First, the evaluation protocol: no train/test split, threshold selection on the evaluation data, and no error bars or significance tests. A 6.6–11% margin could easily shrink or vanish with a proper held-out set or bootstrap intervals. Second, the dataset is small and single-domain (HR handbook, one company). That limits any general claim. Third, the baseline P(yes) is a single SLM on the whole response, so it does not isolate the contribution of the ensemble versus just sentence splitting; the paper attributes the gain to multiple models, but the comparison is confounded. Fourth, no code or data release, so the result can't be checked. The distributions in Section V-F are nice descriptive evidence, but they don't fix the in-sample tuning.

Who gets value: someone building a lightweight hallucination checker for internal RAG applications might find the recipe useful as a starting point, but should not trust the reported numbers. The paper deserves a serious referee—it's coherent, the idea is sensible, and the flaw is fixable with a proper split—but if the evaluation isn't redone, the result should not be published as a 10% improvement.

Referee Report

4 major / 7 minor

Summary. The paper proposes a hallucination-detection framework for RAG-generated answers. It splits each answer into sentences, prompts two small language models (Qwen2 and MiniCPM) to output P(token1=Yes) for each sentence given the question and retrieved context, normalizes these probabilities per model using Eq. (4), averages them across models, and aggregates the sentence scores via a harmonic mean (Eqs. (5)-(6)). A response is classified as correct if the final score exceeds a threshold. The framework is evaluated on a proprietary HR-policy dataset of 'over 100' question-context-response triples with correct, partial, and wrong answers, and the paper reports F1/precision/recall numbers, claiming a 10% F1 improvement over baselines and specifically 11% over ChatGPT and 6.6% over P(yes) in the partial-response condition.

Significance. The idea is practically attractive: local small models avoid API costs, sentence-level decomposition addresses partially correct answers, and ensembling two SLMs is a simple, clearly specified procedure. The equations are easy to follow, and the comparison with P(yes) and ChatGPT is a useful sanity check. However, the empirical support as presented is not valid evidence for the claim: thresholds, normalization moments, and the aggregation rule are all selected on the same data used to report F1, no held-out evaluation or uncertainty quantification is given, and the dataset is small, single-domain, and unreleased. These issues are load-bearing because the headline numbers are in-sample optima. If a re-analysis with proper validation preserves the margin, the contribution would be a useful practical tool, but the current manuscript does not establish that.

major comments (4)

[Section V-D] The reported headline results are in-sample optimized. The text states that the thresholds yielding the highest F1 scores were selected (Fig. 3), but no train/validation/test split is described anywhere in Section V. Because the same data are used both to choose the decision threshold and to compute F1, the reported F1 values (e.g., 0.81 vs 0.73/0.76 in the partial condition) are upper bounds of what the method can achieve after tuning, not estimates of predictive performance. This directly undermines the abstract's 10% improvement claim and the conclusion's 11% and 6.6% claims. The authors should either use a held-out test set with the threshold chosen on a development set, or use nested cross-validation, and should report the threshold values used.
[Section V-C/V-E] The same evaluation-fitting problem applies to the normalization moments in Eq. (4) and to the choice of aggregation mean. Eq. (4) says the moments can be computed based on previous responses, but the paper does not state what was actually done; Fig. 5 compares five aggregation options and selects the harmonic mean because it yields the best outcomes. Thus the per-model normalization moments, the decision threshold, and the aggregation function are all selected on the evaluation data. Without an independent or preregistered choice for these components, the comparison against P(yes) and ChatGPT is not a fair out-of-sample comparison.
[Section V-A/V-D] No uncertainty quantification is provided. Section V-A reports over 100 question/context/response sets, but Figs. 3-5 report single F1/precision/recall values without error bars, confidence intervals, or significance tests. With this sample size, the reported gaps (e.g., 0.81 vs 0.76 for F1 in the partial condition) could easily be within sampling noise. The authors should provide bootstrap confidence intervals or paired tests (e.g., McNemar) on the classifications, and should report the exact dataset size and per-class counts.
[Section IV-B/V-C] The method assumes that P(token1=yes) from the two SLMs is a meaningful correctness signal, but no calibration analysis is provided. Since the decision threshold is selected on the test data, the reported F1 could be an artifact of arbitrary threshold choice even if the scores carry no calibrated probability information. The authors should report calibration curves or at least give exact score distributions with thresholds marked, and show that the chosen threshold is stable across splits or bootstrap resamples.

minor comments (7)

[Eq. (4)] The equation uses the symbols µm and σm, but the surrounding text refers to µj and σj; please unify the notation and clarify whether σ is the standard deviation (the equation uses a z-score, so it should be).
[Eqs. (6) and (8)] The harmonic and geometric means require si,j > 0, but normalized scores in Eq. (4) can be negative or zero; the manuscript states that such values are adjusted but never specifies the adjustment. This matters because the final scores depend on the exact transformation.
[Section V-A] The abstract describes responses with fully and partially correct sentences, but Section V-A says the labels are not applied at the sentence level; please reconcile this wording.
[Section II] ROUGE is cited to reference [21], which is the Guo et al. calibration paper; the citation numbering appears to be off, as [2] already cites Lin's ROUGE paper.
[Section IV-A/V-C] The exact SLM prompt template, decoding parameters (temperature, max new tokens), and the SpaCy sentence-splitting settings are not given; please add these details for reproducibility.
[Figs. 6-7] The histograms are shown without bin counts or exact axis values; the qualitative interpretation would be stronger with the underlying distributions or tables.
[Throughout] There are minor typos (e.g., 'vectoried' in Section I, 'differ ent' in Section III-B) and the abstract's 10% improvement is not qualified as relative or absolute; please state the comparison condition and whether the improvement is relative or absolute.

Circularity Check

2 steps flagged · score 6.0 of 10

Reported F1 gains are in-sample optima: thresholds and the harmonic-mean aggregation rule are selected on the same evaluation data used to report the 10% improvement.

fitted input called prediction [Section V-D, Eqs. 4-6 and Conclusion]
"The results are shown in Fig. 3. Fig. 3 (a) shows the results of detecting “correct” responses from “wrong” responses, while Fig. 3 (b) illustrates detecting “correct” responses from “partial” responses, with the thresholds yielding the highest F 1 scores selected."

The decision threshold is the final classifier parameter that converts the score of Eq. 6 into a “correct” vs. “not correct” label. Selecting this threshold to maximize F1 on the exact responses that are then used to report F1 means every reported F1 value is an in-sample maximum over the threshold, not an estimate of performance on new responses. No train/development/test split is described anywhere in Section V, so the abstract’s “10% improvement in F1 scores” and the conclusion’s “11% and 6.6% better” are fitted, not predicted, results. Eq.
fitted input called prediction [Section V-E, Eq. 6 and Fig. 5]
"Hence, Fig. 5 shows the results of testing different mean calculations. In Fig. 5 (a), the F 1 scores range from 0.75 to 0.99, with the highest score observed for the “max” method (0.99). Fig. 5 (b) shows lower F 1 scores, with the highest score of 0.81 for the “harmonic” method and the lowest at 0.66 for the “min” method. It is observed that the “max” method does not work well for “partial” responses, while there are good correct and hallucination sentences in one response. The results indicate that the harmonic mean yields the best outcomes."

The aggregation rule in Eq. 6 is presented as the final scoring method, but Section V-E shows that it was selected only after comparing arithmetic, geometric, max, min, and harmonic means on the same evaluation data and keeping the one with the best F1. This is an additional model-selection step performed on the test set. Combined with threshold selection in Section V-D, the proposed method’s reported advantage over ChatGPT and P(yes) is the result of optimizing at least two free choices (threshold and aggregation function) against the very labels used for evaluation. The paper does not report whether the same choices transfer to a held-out split, so the claimed 10% improvement is an in-sample optimum rather than a predictive result.

full rationale

The internal derivation of the hallucination score is not definitionally circular: Eqs. 2-6 define the score as a function of SLM output probabilities and sentence splits, not as a restatement of the ground-truth labels, and no imported uniqueness theorem or load-bearing self-citation is used. The only self-citation ([10], in the related-work list) is not load-bearing. However, the central empirical claim--that the proposed framework improves F1 by about 10% over baselines--rests on evaluation-set fitting. Section V-D explicitly says thresholds are chosen as the ones yielding the highest F1 scores on the data being reported, and Section V-E chooses the harmonic mean after comparing several aggregation functions on the same data. Eq. 4 says normalization moments “can be computed based on previous responses,” yet no train/test split or separate moment-estimation procedure is described. With only “over 100” question/context/response sets, tuning a threshold and an aggregation rule against the evaluation labels can substantially inflate or even create the reported margin. Thus the reported performance is an in-sample optimum, not a predictive estimate, which is a partial circularity in the paper’s central claim; the method itself still has independent conceptual content. Score 6 reflects one or more “predictions” that reduce by construction to fitted choices, while stopping short of saying the scoring equations themselves are tautological.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central scoring equations (Eqs. 1 to 6) are simple arithmetic and do not introduce hidden constants. The load-bearing choices are assuming yes-token probability reflects correctness for small models, assuming the retrieved context is the correct reference, assuming sentence splitting preserves all relevant information, and assuming the hand-built labels are reliable. The experimental protocol also has two fitted quantities, the normalization moments and the decision threshold, that appear to be estimated from the evaluation data, plus a post-hoc chosen aggregation rule.

free parameters (3)

Per-model normalization moments (mu_m, sigma_m) = Not reported
Used in Eq. 4 to standardize scores from each SLM. The paper says these can be computed based on previous responses, but no train/test split is described, so the moments may be estimated from the evaluation set, leaking distribution information.
Decision threshold tau = Not reported, chosen to maximize F1
In Section V-D, the threshold for s_i in Eq. 6 is selected as the one yielding the highest F1 on the same data, making the reported F1 a fitted value.
Aggregation mean type = harmonic mean
Eq. 6 uses harmonic mean; Section V-E tests geometric, arithmetic, min, and max and reports harmonic mean yields the best F1. The aggregation family is chosen post hoc on the evaluation data.

assumptions (4)

domain assumption The probability of generating "Yes" as the first token is a faithful measure of the correctness of the checked sentence, and the P(True) property from large models transfers to small models.
Eqs. 1 to 3 treat s_i^(m) = P(token1=yes) as the score; the paper cites Kadavath et al. for large models and asserts even a relatively small model can produce accurate results (Section I) without measuring calibration.
domain assumption The provided context retrieved from the vector database is a complete and correct reference, and the SLM judges sentence accuracy only against this context.
Section III-B and Eq. 3 condition on c_i; no handling of missing or wrong context is described.
domain assumption Splitting the response into sentences with SpaCy preserves all information needed to judge correctness, and each sentence is judged independently.
Section IV-A states splitting avoids confusion, but cross-sentence dependencies and referential expressions across sentences are not handled.
domain assumption The Lane Crawford handbook dataset labels (correct, partial, wrong) are reliable ground truth for evaluation.
Section V-A describes hand-constructed responses without inter-annotator agreement or external validation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Hallucination Detection with Small Language Models." pith.science (2026). https://pith.science/paper/54PHBTB5

@misc{pith2026250622486,
  author       = {Pith},
  title        = {Pith review of: Hallucination Detection with Small Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/54PHBTB5}},
  note         = {Machine review of arXiv:2506.22486}
}

read the original abstract

Since the introduction of ChatGPT, large language models (LLMs) have demonstrated significant utility in various tasks, such as answering questions through retrieval-augmented generation. Context can be retrieved using a vectorized database, serving as a foundation for LLMs to generate responses. However, hallucinations in responses can undermine the reliability of LLMs in practical applications, and they are not easily detectable in the absence of ground truth, particularly in question-and-answer scenarios. This paper proposes a framework that integrates multiple small language models to verify responses generated by LLMs using the retrieved context from a vectorized database. By breaking down the responses into individual sentences and utilizing the probability of generating "Yes" tokens from the outputs of multiple models for a given set of questions, responses, and relevant context, hallucinations can be detected. The proposed framework is validated through experiments with real datasets comprising over 100 sets of questions, answers, and contexts, including responses with fully and partially correct sentences. The results demonstrate a 10\% improvement in F1 scores for detecting correct responses compared to hallucinations, indicating that multiple small language models can be effectively employed for answer verification, providing a scalable and efficient solution for both academic and practical applications.

Figures

Figures reproduced from arXiv: 2506.22486 by the authors.

**Figure 2.** Flow of using the proposed framework: (a) LLMs generate responses; (b) the proposed framework. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Experiment results on the best F1 for different approaches in detecting correct responses from: (a) wrong, (b) partial [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Experiment results on the best p and r for detecting correct responses from: (a) wrong, (b) partial. The incorrect parts are italicized. The reason the response is labeled as partial is that while the working hours are accurate, the days are not. Each question in the d…

**Figure 5.** Figure 5: Experiment results on different means algorithms in detecting correct responses from: (a) wrong, (b) partial. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Distributions of approaches: (a) proposed, (b) [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

**Figure 7.** Figure 7: Distributions of two means: (a) geometric, (b) harmonic. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 6 canonical work pages

[1]

A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions,

L. Huang, W. Yu, W. Ma, W. Zhong, Z. Feng, H. Wang, Q. Chen, W. Peng, X. Feng, B. Qin et al. , “A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions,” arXiv preprint arXiv:2311.05232 , 2023

arXiv 2023
[2]

Rouge: A package for automatic evaluation of summaries,

C.-Y . Lin, “Rouge: A package for automatic evaluation of summaries,” in Text summarization branches out , 2004, pp. 74–81

2004
[3]

Language models (mostly) know what they know,

S. Kadavath, T. Conerly, A. Askell, T. Henighan, D. Drain, E. Perez, N. Schiefer, Z. Hatfield-Dodds, N. DasSarma, E. Tran-Johnson et al. , “Language models (mostly) know what they know,” arXiv preprint arXiv:2207.05221, 2022

arXiv 2022
[4]

Retrieval- augmented generation for knowledge-intensive nlp tasks,

P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. K ¨uttler, M. Lewis, W.-t. Yih, T. Rockt ¨aschel et al. , “Retrieval- augmented generation for knowledge-intensive nlp tasks,” Advances in Neural Information Processing Systems , vol. 33, pp. 9459–9474, 2020

2020
[5]

When large language models meet vector databases: A survey,

Z. Jing, Y . Su, Y . Han, B. Yuan, H. Xu, C. Liu, K. Chen, and M. Zhang, “When large language models meet vector databases: A survey,” arXiv preprint arXiv:2402.01763, 2024

arXiv 2024
[6]

Small language models: Survey, measurements, and insights,

Z. Lu, X. Li, D. Cai, R. Yi, F. Liu, X. Zhang, N. D. Lane, and M. Xu, “Small language models: Survey, measurements, and insights,” arXiv preprint arXiv:2409.15790, 2024

arXiv 2024
[7]

Prompt programming for large language models: Beyond the few-shot paradigm,

L. Reynolds and K. McDonell, “Prompt programming for large language models: Beyond the few-shot paradigm,” in Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems , 2021, pp. 1–7

2021
[8]

Challenges and applications of large language models,

J. Kaddour, J. Harris, M. Mozes, H. Bradley, R. Raileanu, and R. McHardy, “Challenges and applications of large language models,” arXiv preprint arXiv:2307.10169 , 2023

arXiv 2023

Show all 38 references

[9]

Translating natural language to planning goals with large-language models,

Y . Xie, C. Yu, T. Zhu, J. Bai, Z. Gong, and H. Soh, “Translating natural language to planning goals with large-language models,” arXiv preprint arXiv:2302.05128, 2023

2023 arXiv
[10]

A reality check of the benefits of llm in business,

M. Cheung, “A reality check of the benefits of llm in business,” arXiv preprint arXiv:2406.10249, 2024

2024 arXiv
[11]

Recent advances in recurrent neural networks,

H. Salehinejad, S. Sankar, J. Barfett, E. Colak, and S. Valaee, “Recent advances in recurrent neural networks,” arXiv preprint arXiv:1801.01078, 2017

2017 arXiv
[12]

Long short-term memory,

A. Graves and A. Graves, “Long short-term memory,” Supervised sequence labelling with recurrent neural networks , pp. 37–45, 2012

2012
[13]

Overview of the transformer-based models for nlp tasks,

A. Gillioz, J. Casas, E. Mugellini, and O. Abou Khaled, “Overview of the transformer-based models for nlp tasks,” in 2020 15th Conference on Computer Science and Information Systems (FedCSIS) . IEEE, 2020, pp. 179–183

2020
[14]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems , vol. 30, 2017

2017
[15]

Gpt-3: What’s it good for?

R. Dale, “Gpt-3: What’s it good for?” Natural Language Engineering , vol. 27, no. 1, pp. 113–118, 2021

2021
[16]

An overview of bard: an early experiment with generative ai,

J. Manyika, “An overview of bard: an early experiment with generative ai,” Technical report, Google AI, Tech. Rep., 2023

2023
[17]

Evaluating verifiability in generative search engines,

N. F. Liu, T. Zhang, and P. Liang, “Evaluating verifiability in generative search engines,” arXiv preprint arXiv:2304.09848 , 2023

2023 arXiv
[18]

Training language models to follow instructions with human feedback,

L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray et al. , “Training language models to follow instructions with human feedback,” Advances in neural information processing systems , vol. 35, pp. 27 730–27 744, 2022

2022
[19]

Language mod- els are few-shot learners,

T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell et al., “Language mod- els are few-shot learners,” Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020

1901
[20]

” confidently nonsensical?

P. N. Venkit, T. Chakravorti, V . Gupta, H. Biggs, M. Srinath, K. Goswami, S. Rajtmajer, and S. Wilson, “” confidently nonsensical?”: A critical survey on the perspectives and challenges of’hallucinations’ in nlp,” arXiv preprint arXiv:2404.07461 , 2024

2024 arXiv
[21]

On calibration of modern neural networks,

C. Guo, G. Pleiss, Y . Sun, and K. Q. Weinberger, “On calibration of modern neural networks,” in International conference on machine learning. PMLR, 2017, pp. 1321–1330

2017
[22]

Bleu: a method for automatic evaluation of machine translation,

K. Papineni, S. Roukos, T. Ward, and W.-J. Zhu, “Bleu: a method for automatic evaluation of machine translation,” in Proceedings of the 40th annual meeting of the Association for Computational Linguistics , 2002, pp. 311–318

2002
[23]

Hallucination detection: Robustly discerning reliable answers in large language models,

Y . Chen, Q. Fu, Y . Yuan, Z. Wen, G. Fan, D. Liu, D. Zhang, Z. Li, and Y . Xiao, “Hallucination detection: Robustly discerning reliable answers in large language models,” in Proceedings of the 32nd ACM International Conference on Information and Knowledge Management , 2023, p...

2023
[24]

Controlled hallucinations: Learning to generate faithfully from noisy data,

K. Filippova, “Controlled hallucinations: Learning to generate faithfully from noisy data,” arXiv preprint arXiv:2010.05873 , 2020

2010 arXiv
[25]

Llms know more than they show: On the intrinsic representation of llm hallucinations,

H. Orgad, M. Toker, Z. Gekhman, R. Reichart, I. Szpektor, H. Kotek, and Y . Belinkov, “Llms know more than they show: On the intrinsic representation of llm hallucinations,” arXiv preprint arXiv:2410.02707 , 2024

2024 arXiv
[26]

Glore: When, where, and how to improve llm reasoning via global and local refinements,

A. Havrilla, S. Raparthy, C. Nalmpantis, J. Dwivedi-Yu, M. Zhuravin- skyi, E. Hambro, and R. Railneau, “Glore: When, where, and how to improve llm reasoning via global and local refinements,” arXiv preprint arXiv:2402.10963, 2024

2024 arXiv
[27]

Generating sequences by learning to self-correct,

S. Welleck, X. Lu, P. West, F. Brahman, T. Shen, D. Khashabi, and Y . Choi, “Generating sequences by learning to self-correct,” arXiv preprint arXiv:2211.00053, 2022

2022 arXiv
[28]

Detecting hallucinations in large language models using semantic entropy,

S. Farquhar, J. Kossen, L. Kuhn, and Y . Gal, “Detecting hallucinations in large language models using semantic entropy,” Nature, vol. 630, no. 8017, pp. 625–630, 2024

2024
[29]

To believe or not to believe your llm,

Y . A. Yadkori, I. Kuzborskij, A. Gy ¨orgy, and C. Szepesv´ari, “To believe or not to believe your llm,” arXiv preprint arXiv:2406.02543 , 2024

2024 arXiv
[30]

To cot or not to cot? chain-of-thought helps mainly on math and symbolic reasoning,

Z. Sprague, F. Yin, J. D. Rodriguez, D. Jiang, M. Wadhwa, P. Singhal, X. Zhao, X. Ye, K. Mahowald, and G. Durrett, “To cot or not to cot? chain-of-thought helps mainly on math and symbolic reasoning,” arXiv preprint arXiv:2409.12183, 2024

2024 arXiv
[31]

Improving language understanding by generative pre-training,

A. Radford, K. Narasimhan, T. Salimans, I. Sutskever et al., “Improving language understanding by generative pre-training,” 2018

2018
[32]

(accessed: 12.11.2023)

(2023) Llama-2-70b. (accessed: 12.11.2023). [Online]. Available: https://huggingface.co/meta-llama/Llama-2-70b

2023
[33]

Lora: Low-rank adaptation of large language models,

E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “Lora: Low-rank adaptation of large language models,” arXiv preprint arXiv:2106.09685 , 2021

2021 arXiv
[34]

Reducing hallucination in struc- tured outputs via retrieval-augmented generation,

P. B ´echard and O. M. Ayala, “Reducing hallucination in struc- tured outputs via retrieval-augmented generation,” arXiv preprint arXiv:2404.08189, 2024

2024 arXiv
[35]

Qwen technical report,

J. Bai, S. Bai, Y . Chu, Z. Cui, K. Dang, X. Deng, Y . Fan, W. Ge, Y . Han, F. Huang et al. , “Qwen technical report,” arXiv preprint arXiv:2309.16609, 2023

2023 arXiv
[36]

Minicpm: Unveiling the potential of small language models with scalable training strategies,

S. Hu, Y . Tu, X. Han, C. He, G. Cui, X. Long, Z. Zheng, Y . Fang, Y . Huang, W. Zhao et al. , “Minicpm: Unveiling the potential of small language models with scalable training strategies,” arXiv preprint arXiv:2404.06395, 2024

2024 arXiv
[37]

Mixture-of-experts with expert choice routing,

Y . Zhou, T. Lei, H. Liu, N. Du, Y . Huang, V . Zhao, A. M. Dai, Q. V . Le, J. Laudon et al. , “Mixture-of-experts with expert choice routing,” Advances in Neural Information Processing Systems , vol. 35, pp. 7103– 7114, 2022

2022
[38]

Complex claim verification with evidence retrieved in the wild,

J. Chen, G. Kim, A. Sriram, G. Durrett, and E. Choi, “Complex claim verification with evidence retrieved in the wild,” arXiv preprint arXiv:2305.11859, 2023

2023 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions,

L. Huang, W. Yu, W. Ma, W. Zhong, Z. Feng, H. Wang, Q. Chen, W. Peng, X. Feng, B. Qin et al. , “A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions,” arXiv preprint arXiv:2311.05232 , 2023

arXiv 2023

[2] [2]

Rouge: A package for automatic evaluation of summaries,

C.-Y . Lin, “Rouge: A package for automatic evaluation of summaries,” in Text summarization branches out , 2004, pp. 74–81

2004

[3] [3]

Language models (mostly) know what they know,

S. Kadavath, T. Conerly, A. Askell, T. Henighan, D. Drain, E. Perez, N. Schiefer, Z. Hatfield-Dodds, N. DasSarma, E. Tran-Johnson et al. , “Language models (mostly) know what they know,” arXiv preprint arXiv:2207.05221, 2022

arXiv 2022

[4] [4]

Retrieval- augmented generation for knowledge-intensive nlp tasks,

P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. K ¨uttler, M. Lewis, W.-t. Yih, T. Rockt ¨aschel et al. , “Retrieval- augmented generation for knowledge-intensive nlp tasks,” Advances in Neural Information Processing Systems , vol. 33, pp. 9459–9474, 2020

2020

[5] [5]

When large language models meet vector databases: A survey,

Z. Jing, Y . Su, Y . Han, B. Yuan, H. Xu, C. Liu, K. Chen, and M. Zhang, “When large language models meet vector databases: A survey,” arXiv preprint arXiv:2402.01763, 2024

arXiv 2024

[6] [6]

Small language models: Survey, measurements, and insights,

Z. Lu, X. Li, D. Cai, R. Yi, F. Liu, X. Zhang, N. D. Lane, and M. Xu, “Small language models: Survey, measurements, and insights,” arXiv preprint arXiv:2409.15790, 2024

arXiv 2024

[7] [7]

Prompt programming for large language models: Beyond the few-shot paradigm,

L. Reynolds and K. McDonell, “Prompt programming for large language models: Beyond the few-shot paradigm,” in Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems , 2021, pp. 1–7

2021

[8] [8]

Challenges and applications of large language models,

J. Kaddour, J. Harris, M. Mozes, H. Bradley, R. Raileanu, and R. McHardy, “Challenges and applications of large language models,” arXiv preprint arXiv:2307.10169 , 2023

arXiv 2023