Pith. sign in

REVIEW 4 major objections 6 minor 28 references

Cleanse: Uncertainty Estimation Approach Using Clustering-based Semantic Consistency in LLMs

T0 review · 4 major / 6 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 Cleanse measures uncertainty as the proportion of intra-cluster similarity in total similarity, and the paper reports that this ratio detects LLM hallucinations better than token-level baselines.

desk verdict Cleanse proposes a genuinely new ratio-of-similarities uncertainty score, but the paper omits the one baseline that matters most—semantic entropy—leaving the headline claim unproven. read the letter →

arxiv 2507.14649 v1 pith:JHN53VR2 submitted 2025-07-19 cs.CL cs.AI

Minsuh Joo , Hyunsoo Cho This is my paper
classification cs.CLcs.AI
keywords uncertaintyestimationhallucinationdetectionsemanticconsistencyclusteringnaturallanguageinferencehiddenembeddingslargemodelsquestionanswering
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 proposes Cleanse, an uncertainty score that measures how semantically consistent a set of generated answers is. It clusters repeated generations by bi-directional NLI entailment, extracts hidden embeddings from the middle layer, and sets the score to the fraction of total cosine similarity that lies inside clusters. The central claim is that this ratio separates correct from hallucinated answers better than token-level perplexity, entropy, lexical similarity, and a plain average cosine score. If true, Cleanse gives a practical white-box signal for when to distrust an LLM output without retraining or external knowledge.

What carries the argument

The load-bearing mechanism is bi-directional NLI entailment clustering of the sampled outputs, combined with the last-token hidden embedding from the middle layer of the LLM as the semantic representation. A pair of outputs joins the same cluster only if each entails the other when concatenated with the question, and transitivity lets a new output join a cluster by entailing one existing member. The Cleanse Score then converts the cluster structure into a single ratio: the sum of cosine similarities within clusters divided by the sum over all pairs, so inter-cluster similarity acts as an explicit penalty rather than being averaged away.

What would settle it

Replace the NLI clusterer with random cluster assignments of the same cluster sizes and rerun the SQuAD and CoQA evaluations; if the random clusters preserve most of the AUROC gain over the cosine baseline, then the semantic-consistency mechanism is not doing the work. Alternatively, test on a set of paraphrases with negation or coreference shifts where a strong NLI model is known to misjudge entailment: if Cleanse assigns high scores to clearly wrong answers or low scores to clearly correct ones at rates no better than the cosine baseline, the claimed advantage of the inter-cluster penalty would fail.

Watch

Extended reading notes

Core claim

Cleanse quantifies hallucination by treating the proportion of intra-cluster similarity in total similarity as the degree of consistency among sampled generations. The score is defined as Cleanse Score = intra-cluster sim. / total sim., where intra-cluster similarity sums cosine similarities between hidden embeddings inside the same semantic cluster and total similarity sums all pairwise cosine similarities. High scores mean the generations fall into few semantically equivalent clusters with dense internal agreement; low scores mean outputs scatter into many clusters with substantial cross-cluster similarity, which the method counts as a penalty. The paper reports that Cleanse achieves the highest AUROC among all baselines for LLaMA-7B, LLaMA-13B, LLaMA2-7B, and Mistral-7B on SQuAD and CoQA, with the largest margins over lexical similarity on Mistral-7B, and that its advantage grows as the Rouge-L correctness threshold becomes stricter.

Load-bearing premise

The whole score depends on the NLI-based clustering correctly grouping answers that mean the same thing; if the NLI model puts equivalent answers in different clusters or different answers in the same cluster, intra-cluster similarity no longer tracks semantic consistency and the score stops separating correct from hallucinated outputs.

Editorial extensions

If this is right

  • Cleanse provides an uncertainty score for any white-box LLM without additional training: it needs only multiple sampled generations, hidden embeddings, and a small NLI model for clustering.
  • Because the score is a ratio with inter-cluster similarity as a penalty, divergent outputs lower the score even when average pairwise similarity is high, which addresses a known failure mode of sentence-level averaging.
  • The reported AUROC gains at stricter Rouge-L thresholds suggest Cleanse is most useful in tasks with exact correctness criteria, such as question answering and translation.
  • The clustering-model comparison indicates that the choice of NLI model matters, and that a clusterer producing few clusters for correct answers and many for incorrect answers yields better hallucination detection.

Reading between the lines

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

  • Because Cleanse relies on hidden embeddings, it could likely be adapted to other layers or to embeddings produced by the NLI model itself, which would extend it to black-box LLMs; the paper does not test these variants.
  • The score weights every pair equally inside a cluster; weighting by cluster size or using cluster-level representatives instead of pairwise sums might sharpen the separation between confident and uncertain answers beyond what the paper reports.
  • Cleanse measures semantic consistency, not factual correctness; a model that is confidently and consistently wrong would still receive a high score, so the method is best used as a complement to correctness-oriented checks rather than a substitute.
  • The reported threshold experiments suggest a testable extension: measuring whether Cleanse remains advantageous when the number of sampled generations K is varied, since the paper fixes K implicitly in its experiments.
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

4 major / 6 minor

Summary. The paper proposes Cleanse, a white-box uncertainty estimation method for detecting hallucinations in LLM question answering. For each question, K generations are produced; the generations are clustered into semantic equivalence classes using bidirectional NLI entailment with question-answer concatenation, following Kuhn et al. (2023). A hidden embedding is extracted for each generation, and the Cleanse Score is defined in Eq. (3) as the ratio of intra-cluster cosine similarity to total pairwise cosine similarity across all generations. The method is evaluated on SQuAD and CoQA with LLaMA-7B, LLaMA-13B, LLaMA2-7B, and Mistral-7B, using AUROC and PCC to separate correct from incorrect generations, against perplexity, LN-entropy, lexical similarity, and a simple cosine-score baseline. The authors claim Cleanse consistently outperforms all baselines and remains robust under stricter Rouge-L correctness thresholds.

Significance. If the central claim holds, Cleanse is a practical and attractively simple uncertainty signal: Eq. (3) contains no fitted constants, the clustering step is already established in the literature, and the reported AUROC improvements are consistent across eight model/dataset settings, especially for Mistral-7B. The threshold-sweep experiments in Table 3 and the clustering-model comparison in Table 2 are useful sensitivity analyses. However, the paper's contribution is defined relative to existing uncertainty baselines, and the most directly related method—semantic entropy, which uses the very same NLI clusters—is not evaluated. In addition, the clustering model is selected on the test set, and no error bars or significance tests are provided. These issues leave the headline claim of consistent superiority insufficiently supported, although the method itself appears defensible and worth revising.

major comments (4)
  1. [Section 2, Section 3.2, Table 1] The paper cites semantic entropy (Kuhn et al., 2023) and adopts its bidirectional NLI clustering algorithm, but semantic entropy is not included as a baseline in Table 1. Since Cleanse's only difference from semantic entropy is the scoring function over the same clusters—entropy over cluster distribution versus the ratio in Eq. (3)—the semantic-entropy baseline is the natural and most informative comparison. Without it, the reader cannot tell whether the ratio of intra-cluster to total cosine similarity is an actual improvement over the standard use of these clusters. I ask the authors to add semantic entropy (and ideally SAR, which is cited in Section 2) to Table 1 using the same number of sampled outputs and the same clustering assignments, and to report the AUROC/PCC differences explicitly.
  2. [Section 4.2, Table 1] The text states that Cleanse Score outperforms all four baselines 'when evaluated using AUROC and PCC,' but Table 1 shows two settings where Cosine Score has higher PCC than Cleanse Score: LLaMA-7B on CoQA (48.4 vs 47.6) and LLaMA2-7B on CoQA (50.6 vs 49.7). The claim should be corrected to 'higher AUROC in all settings and higher PCC in most settings,' and the two PCC failures should be discussed—for example, whether they are within the expected variation given that no error bars or significance tests are reported.
  3. [Section 4.2, Table 2] The clustering model nli-deberta-v3-base is selected by comparing AUROC on the same test sets used for the main evaluation, and the auxiliary criterion (difference in average number of clusters for correct versus incorrect generations) is also computed on test data. This is selection on the test set, which makes the reported Cleanse AUROC values optimistic for the chosen NLI model. The authors should either perform model selection on a held-out development set or report the main results averaged over the four clustering models, or otherwise account for the selection effect.
  4. [Section 4.1, Sections 3.1 and 3.3] Several experimental parameters essential for reproducibility are not reported: the number of sampled outputs K, the decoding temperature and sampling strategy, the exact hidden layer used (the paper only says 'middle layer'), and the precise implementation of the Rouge-L threshold (e.g., whether F1 or recall is used). In addition, no error bars, confidence intervals, or significance tests are given, even though some AUROC differences in Table 1 are small (for example, 79.4 vs 78.5 for LLaMA-7B on CoQA). These omissions make it difficult to determine whether the reported gains are reliable.
minor comments (6)
  1. [Section 4.1] There is a typo in the baselines paragraph: 'baeslines' should be 'baselines'.
  2. [Section 4.1] The parenthetical 'SQuAD (20.92) has longer ground truth answer spans than CoQA (13.67)' is unclear: the reader is not told what 20.92 and 13.67 represent (presumably average answer length in tokens or characters), nor how they are computed.
  3. [Figure 5 caption] The caption says 'having few clusters for correct answers and a few clusters for wrong answers is advantageous,' which contradicts the method's own rationale that incorrect generations should produce many clusters. This should read 'few clusters for correct answers and many clusters for wrong answers.'
  4. [Algorithm 1] The line 's(c) ← c0' is informal and unclear; it should be described as selecting a representative element of the existing cluster c for comparison. The pseudocode would also benefit from specifying the order of comparisons and the early-exit condition for transitivity.
  5. [Section 3.3] Eq. (1) defines intra-cluster similarity, but inter-cluster similarity is only described in words. For clarity, an explicit formula for inter-cluster similarity should be added before Eq. (3).
  6. [Section 4.2] The discussion of Figure 4 refers to 'the performance of lexical similarity either remains the same or decreases as the Rouge-L threshold increases,' but this statement depends on the model/dataset setting; Table 3 shows an increase for Mistral-7B on SQuAD (69.0 to 70.7). The sentence should be rephrased to reflect the mixed pattern.

Circularity Check

1 steps flagged · score 2.0 of 10

Cleanse score itself is not circular, but the clustering model is selected on the evaluation AUROC before the headline comparison, a mild fitted-input issue.

  1. fitted input called prediction [Section 4.2, Clustering model comparison (Table 2) and Table 1 setup]
    "We compare four fine-tuned NLI model, deberta-large-mnli (He et al., 2020), roberta-large-mnli (Liu et al., 2019), nli-deberta-v3-base (He et al., 2021) and nli-deberta-v3-large (He et al., 2021) to find the optimal clustering model. ... Table 2 shows that AUROC scores of Cleanse Score using nli-deberta-v3-base are slightly better than when using other clustering models."

    The choice of the clustering model is made by maximizing AUROC on the same SQuAD/CoQA test sets that Table 1 uses to claim Cleanse consistently outperforms baselines. Therefore the reported Table 1 performance of Cleanse is not an out-of-sample prediction for the selected configuration; the 'nli-deberta-v3-base is best' result is the selection criterion itself. This inflates the headline AUROC comparison, though the Cleanse Score formula (Eq. 3) itself contains no fitted parameters and is not derived from the correctness labels.

full rationale

The central derivation of Cleanse is self-contained: Eq. 3 defines the score as intra-cluster cosine similarity divided by total similarity, using hidden embeddings and bi-directional NLI clusters; no correctness labels or fitted constants enter the score. The empirical claim that Cleanse detects hallucinations better than baselines is therefore not circular by construction. The only significant circularity-adjacent issue is that the NLI clustering model is selected by the same AUROC computed on the evaluation sets (Table 2), so the headline Table 1 results are mildly optimistic and the statement that nli-deberta-v3-base is best is the selection objective rather than a fresh finding. The absence of semantic entropy (Kuhn et al., 2023) as a baseline is a serious experimental omission but is not a circularity: Cleanse uses the same clustering but a different scoring function. Overall, no equation-level reduction of prediction to input occurs, so the paper is mostly self-contained with a mild test-set-selection concern.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The core score needs no fitted numeric constants, but the experimental pipeline relies on several unspecified choices (K, layer, temperature) and one test-set-driven selection (clustering model). The semantic embedding and clustering assumptions are borrowed from prior work and not independently validated here.

free parameters (5)
  • Number of sampled outputs K = not reported
    The score is computed over K generated outputs, and the pairwise sums in Eqs. 1-2 scale with K, but K is never stated in the paper.
  • Hidden layer index for hidden embedding = not reported ('middle layer')
    Section 3.1 specifies the last token embedding from the middle layer, but not which layer index; cosine similarities depend on this choice.
  • Decoding temperature = not reported
    The diversity of sampled outputs, and therefore cluster structure, depends on sampling temperature, which is not given.
  • Rouge-L correctness threshold = 0.7 main, 0.5-0.9 sweep
    Correctness labels are defined by Rouge-L above a threshold; the main results use 0.7, and the robustness analysis varies it.
  • Clustering model selection = nli-deberta-v3-base
    The NLI model was chosen because it gave the highest AUROC on the test sets in Table 2, a form of performance-based selection on evaluation data.
assumptions (5)
  • domain assumption Hidden embeddings from the middle layer capture enough semantic information to support uncertainty estimation.
    Invoked in Section 3.1, based on Azaria and Mitchell (2023), without a layer ablation in this paper.
  • domain assumption Bi-directional NLI entailment yields a meaningful partition of outputs into semantic equivalence classes.
    The clustering algorithm in Section 3.2 is adopted from Kuhn et al. (2023) and is the sole source of cluster assignments used in the score.
  • domain assumption Cosine similarity between hidden embeddings reflects semantic relatedness.
    Eqs. 1-2 use cosine similarity as the consistency measure, with no validation that cosine distance matches semantic distance in this space.
  • domain assumption Repeated generations from the same model are more consistent when the model is confident.
    The foundational premise stated in the introduction and used to interpret the score as uncertainty.
  • domain assumption Rouge-L score relative to a single ground-truth answer is a valid correctness measure for QA generations.
    Section 4.1 defines correct answers by Rouge-L threshold 0.7; no human agreement or alternative correctness measure is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cleanse: Uncertainty Estimation Approach Using Clustering-based Semantic Consistency in LLMs." pith.science (2026). https://pith.science/paper/JHN53VR2

@misc{pith2026250714649,
  author       = {Pith},
  title        = {Pith review of: Cleanse: Uncertainty Estimation Approach Using Clustering-based Semantic Consistency in LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JHN53VR2}},
  note         = {Machine review of arXiv:2507.14649}
}
read the original abstract

Despite the outstanding performance of large language models (LLMs) across various NLP tasks, hallucinations in LLMs--where LLMs generate inaccurate responses--remains as a critical problem as it can be directly connected to a crisis of building safe and reliable LLMs. Uncertainty estimation is primarily used to measure hallucination levels in LLM responses so that correct and incorrect answers can be distinguished clearly. This study proposes an effective uncertainty estimation approach, \textbf{Cl}ust\textbf{e}ring-based sem\textbf{an}tic con\textbf{s}ist\textbf{e}ncy (\textbf{Cleanse}). Cleanse quantifies the uncertainty with the proportion of the intra-cluster consistency in the total consistency between LLM hidden embeddings which contain adequate semantic information of generations, by employing clustering. The effectiveness of Cleanse for detecting hallucination is validated using four off-the-shelf models, LLaMA-7B, LLaMA-13B, LLaMA2-7B and Mistral-7B and two question-answering benchmarks, SQuAD and CoQA.

Figures

Figures reproduced from arXiv: 2507.14649 by the authors.

Figure 1
Figure 1. Illustration of Cleanse pipeline. leveraging semantics of responses by employing sentence-level embeddings and bi-directional clus￾tering. First, we generate multiple outputs and extract their hidden embeddings from the model. Then, we cluster those outputs based on their se￾mantic equivalency. Finally, to assess uncertainty, we compute similarities within and across these clusters respectively and calculate Cleanse… view at source ↗
Figure 2
Figure 2. Each white circle indicates a single hidden em [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Case 1 has a small number of clusters, result [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: AUROC difference between Cleanse Score and lexical similarity across four models on two QA datasets, [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: The illustration that shows the importance of [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 5 canonical work pages

  1. [1]

    Zahid Ansari, Mohammad Fazle Azeem, Waseem Ahmed, and A Vinaya Babu. 2015. Quantitative evaluation of performance and validity indices for clustering the web navigational sessions. arXiv preprint arXiv:1507.03340

    work page Pith review arXiv 2015

  2. [2]

    Amos Azaria and Tom Mitchell. 2023. The internal state of an llm knows when it's lying. arXiv preprint arXiv:2304.13734

    arXiv 2023

  3. [3]

    Chao Chen, Kai Liu, Ze Chen, Yi Gu, Yue Wu, Mingyuan Tao, Zhihang Fu, and Jieping Ye. 2024. Inside: Llms' internal states retain the power of hallucination detection. arXiv preprint arXiv:2402.03744

    arXiv 2024

  4. [4]

    Jinhao Duan, Hao Cheng, Shiqi Wang, Alex Zavalny, Chenan Wang, Renjing Xu, Bhavya Kailkhura, and Kaidi Xu. 2023. Shifting attention to relevance: Towards the predictive uncertainty quantification of free-form large language models. arXiv preprint arXiv:2307.01379

    arXiv 2023

  5. [5]

    Shahul Es, Jithin James, Luis Espinosa Anke, and Steven Schockaert. 2024. Ragas: Automated evaluation of retrieval augmented generation. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics: System Demonstrations, pages 150--158

    work page 2024

  6. [6]

    Xiang Gao, Jiaxin Zhang, Lalla Mouatadid, and Kamalika Das. 2024. Spuq: Perturbation-based uncertainty quantification for large language models. arXiv preprint arXiv:2403.02509

    arXiv 2024

  7. [7]

    Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2021. Debertav3: Improving deberta using electra-style pre-training with gradient-disentangled embedding sharing. arXiv preprint arXiv:2111.09543

    arXiv 2021

  8. [8]

    Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2020. Deberta: Decoding-enhanced bert with disentangled attention. arXiv preprint arXiv:2006.03654

    arXiv 2020

Show all 28 references
  1. [9]

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. 2025. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Info...

    2025

  2. [10]

    Ziwei Ji, Delong Chen, Etsuko Ishii, Samuel Cahyawijaya, Yejin Bang, Bryan Wilie, and Pascale Fung. 2024. Llm internal states reveal hallucination risk faced with a query. arXiv preprint arXiv:2407.03282

    2024 arXiv

  3. [11]

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of hallucination in natural language generation. ACM computing surveys, 55(12):1--38

    2023

  4. [12]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas ...

    2023 arXiv

  5. [13]

    Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. 2023. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. arXiv preprint arXiv:2302.09664

    2023 arXiv

  6. [14]

    Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74--81

    2004

  7. [15]

    Stephanie Lin, Jacob Hilton, and Owain Evans. 2022 a . Teaching models to express their uncertainty in words. arXiv preprint arXiv:2205.14334

    2022 arXiv

  8. [16]

    Zhen Lin, Shubhendu Trivedi, and Jimeng Sun. 2023. Generating with confidence: Uncertainty quantification for black-box large language models. arXiv preprint arXiv:2305.19187

    2023 arXiv

  9. [17]

    Zi Lin, Jeremiah Zhe Liu, and Jingbo Shang. 2022 b . Towards collaborative neural-symbolic graph semantic parsing via uncertainty. Findings of the Association for Computational Linguistics: ACL 2022

    2022

  10. [18]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692

    2019 arXiv

  11. [19]

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

    2020 arXiv

  12. [20]

    Ella Rabinovich, Samuel Ackerman, Orna Raz, Eitan Farchi, and Ateret Anaby-Tavor. 2023. Predicting question-answering performance of large language models through semantic consistency. arXiv preprint arXiv:2311.01152

    2023 arXiv

  13. [21]

    Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. Squad: 100,000+ questions for machine comprehension of text. arXiv preprint arXiv:1606.05250

    2016 arXiv

  14. [22]

    Siva Reddy, Danqi Chen, and Christopher D Manning. 2019. Coqa: A conversational question answering challenge. Transactions of the Association for Computational Linguistics, 7:249--266

    2019

  15. [23]

    Jie Ren, Jiaming Luo, Yao Zhao, Kundan Krishna, Mohammad Saleh, Balaji Lakshminarayanan, and Peter J. Liu. 2023. https://arxiv.org/abs/2209.15558 Out-of-distribution detection and selective generation for conditional language models . Preprint, arXiv:2209.15558

    2023 arXiv

  16. [24]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth \'e e Lacroix, Baptiste Rozi \`e re, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023 a . Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971

    2023 arXiv

  17. [25]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023 b . Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288

    2023 arXiv

  18. [26]

    Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, et al. 2023. Siren’s song in the ai ocean: A survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2(5)

    2023 arXiv

  19. [27]

    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...

  20. [28]

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