Pith. sign in

REVIEW 4 major objections 6 minor 29 references

Evaluating Large Language Models for Phishing Detection, Self-Consistency, Faithfulness, and Explainability

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 Self-consistent LLM explanations do not imply accurate phishing detection.

desk verdict Asking a good question and reporting a striking dissociation, but the table behind the headline says it used baseline models, not the fine-tuned models the abstract is about. read the letter →

arxiv 2506.13746 v1 pith:AHUISNWA submitted 2025-06-16 cs.CR cs.AIcs.LG

Shova Kuikel , Aritran Piplai , Palvi Aggarwal This is my paper
classification cs.CRcs.AIcs.LG
keywords phishingdetectionlargelanguagemodelsself-consistencyfaithfulnessexplainabilityCC-SHAPfine-tuningdirectpreferenceoptimization
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 asks whether large language models can both classify phishing emails accurately and produce explanations that are internally consistent with their predictions. The authors fine-tune BERT, Llama 7B, Llama 8B, and Wizard 7B on a balanced corpus of 2,500 phishing and 2,500 ham emails using binary classification, contrastive learning, and direct preference optimization, then measure each model with CC-SHAP, a score that quantifies token-level alignment between prediction and explanation. The central finding is a tradeoff: Llama models show high CC-SHAP scores (above 0.95) but low phishing accuracy (30–40%), while Wizard 7B shows low CC-SHAP (about 0.12) but 80% phishing accuracy. This suggests that self-consistent explanations in fine-tuned LLMs do not imply reliable phishing classification, and that accuracy can be achieved without explanation consistency. If this holds, evaluation of LLM-based phishing detectors must treat accuracy and explainability as separate axes instead of assuming one follows from the other.

What carries the argument

The central object is CC-SHAP, a consistency measure based on SHAPley values: the model's prediction SHAP vector (per-token contributions to the phishing/ham decision) is computed via Monte Carlo masking, and compared against the explanation SHAP vector (per-token contributions to the generated explanation), with the score defined as one minus the cosine distance between their L1-normalized vectors. It is used to quantify whether the tokens a model says it relies on are the tokens that actually drive its decision. The supporting machinery is the fine-tuning setup: LoRA-based parameter-efficient fine-tuning under three objectives (binary sequence classification, contrastive learning, and direct preference optimization) over the same curated phishing/ham corpus.

What would settle it

Run the same CC-SHAP evaluation on a fresh holdout set of 40 or more phishing and ham emails that were never used in fine-tuning and that were collected after the models' training data cutoff; if Wizard 7B's phishing accuracy on that set falls to near Llama's level while its CC-SHAP remains low, the claimed accuracy–consistency tradeoff disappears and becomes an artifact of the evaluation sample.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that prediction–explanation token alignment, as measured by CC-SHAP, is decoupled from classification accuracy after fine-tuning for phishing detection. Llama models produce explanations whose SHAP-attributed tokens closely match the tokens driving their predictions (CC-SHAP above 0.95), yet they classify only 40% and 30% of phishing emails correctly. Wizard 7B, with CC-SHAP around 0.12 for phishing, correctly identifies 80% of phishing emails. The paper interprets this as evidence that high internal consistency in an LLM's reasoning path says nothing about whether that reasoning path is correct, and that a model can reach accurate decisions through a token-weighting pattern that is not faithfully aligned with its stated explanation.

Load-bearing premise

The claim rests on treating the 20 phishing and 20 ham emails used for CC-SHAP scoring as a representative, non-leaked sample of the fine-tuning distribution; if those emails overlap with training examples or reflect older phishing tactics, the reported accuracy–consistency tradeoff would not generalize.

Editorial extensions

If this is right

  • In phishing detection, reporting model accuracy alone, or explanation consistency alone, is insufficient; both must be measured and reported independently.
  • A high CC-SHAP score should not be treated as a reliability certificate for an LLM-based classifier, since it can coincide with near-chance phishing accuracy.
  • Fine-tuning objectives that optimize explanation consistency are not automatically optimizing detection performance; the same objective can yield strong alignment and weak classification.
  • The near-perfect ham accuracy alongside low phishing accuracy across models points to a possible legitimate-email bias that would need monitoring in real deployments.

Reading between the lines

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

  • The CC-SHAP tradeoff is measured on only 20 phishing and 20 ham emails drawn from the same sources used for fine-tuning; a natural extension is to re-run the measurement on a larger, temporally later, and explicitly held-out set to see whether the accuracy–consistency decoupling persists.
  • If the decoupling holds on truly novel phishing emails, explanation faithfulness should be evaluated not only as internal token alignment but also as alignment with features a human expert would cite, since internal consistency alone may be an unreliable proxy for trustworthy reasoning.
  • Wizard 7B's combination of 80% phishing accuracy and very low CC-SHAP suggests that low-calibration models can still be practically useful for detection; a testable extension is whether its correct choices are driven by robust stylistic cues that the SHAP-based explanation framework fails to surface.
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 fine-tunes BERT, Llama-2-7B, Llama-3-8B, and Wizard-7B on a combined phishing/ham dataset (2,500 Nazario phishing emails and 2,500 Enron ham emails) using three fine-tuning approaches: binary sequence classification, contrastive learning, and direct preference optimization. It reports training and validation losses and accuracies, and then evaluates prediction-explanation token alignment using the CC-SHAP metric on a 40-email subset. The central claim, stated in the abstract and Section 6.2, is that Llama models achieve high CC-SHAP scores (around 0.95) but low phishing accuracy (30–40%), whereas Wizard-7B achieves lower CC-SHAP scores (around 0.12) but higher phishing accuracy (80%), suggesting that self-consistent explanations do not imply accurate phishing classification. The paper concludes that binary classification is the most reliable fine-tuning approach and raises questions about balancing human-like uncertainty with reliability.

Significance. If substantiated, the paper would make a useful contribution to the phishing-detection and LLM-explainability communities by empirically separating self-consistency of explanations from task accuracy, and by warning against interpreting high prediction-explanation alignment as a proxy for decision quality. The comparison of binary classification, contrastive learning, and DPO across multiple architectures is also a valuable reference point for practitioners, and the release of code is a positive step for reproducibility. However, the quantitative support for the headline claim is currently weak: the key accuracy figures in Table 4 rest on only 40 emails without confidence intervals or significance tests, the evaluation subset appears to be drawn from the same corpora used for fine-tuning without a documented holdout, and the wording 'using the baseline models' in Section 6.2 directly conflicts with the abstract's attribution of the results to fine-tuned models. These issues are load-bearing, because the paper's central conclusion depends entirely on the validity and interpretation of Table 4.

major comments (4)
  1. [Section 6.2, Table 4] The phrase 'using the baseline models' directly contradicts the abstract and conclusion, which attribute the CC-SHAP/accuracy trade-off to the fine-tuned models. If 'baseline' means the original pre-trained checkpoints (for example, meta-llama/Llama-2-7b-hf without LoRA adapters), then the 40% and 30% phishing accuracies for Llama models and the 80% for Wizard are not properties of the fine-tuned systems evaluated in Tables 2 and 3, and the central claim has no experimental basis. Please specify exactly which checkpoint and which fine-tuning variant (binary classification, contrastive learning, or DPO) produced each row of Table 4, or rerun the CC-SHAP evaluation on the fine-tuned models and report those results.
  2. [Section 2 and Section 6.2] The evaluation subset for CC-SHAP consists of 20 phishing emails from Nazario and 20 ham emails from Enron, and Section 2 states that 'a subset of the same data was also utilized to evaluate the CC-SHAP method.' No sampling method, random seed, temporal split, or holdout procedure is described, so the subset may overlap with the training data used in Section 4.2. This risks data leakage and makes the reported accuracies and CC-SHAP scores ungeneralizable. Please describe the selection procedure in detail, ensure that the evaluation emails are disjoint from all fine-tuning examples, and report the temporal distribution of the phishing emails relative to the 2015–2024 range used for training.
  3. [Table 4 and Section 6.2] With only 20 phishing emails per model, the reported accuracies (40%, 30%, 80%) have wide confidence intervals; for example, an 80% accuracy on 20 items has a 95% confidence interval of roughly 56–94%, and a 40% accuracy has an interval of roughly 19–64%. The paper reports mean ± standard deviation for CC-SHAP scores but no uncertainty for the accuracy percentages. Please include confidence intervals or exact binomial tests for the accuracy figures, and avoid presenting the accuracy differences as conclusive without such quantification.
  4. [Section 5, equations (1)–(3)] The explanation generation procedure is not specified. The paper does not state what prompt template was used to elicit natural language explanations, what decoding parameters (temperature, max tokens, sampling) were used, or how the generated explanation was tokenized and aligned with the input email tokens for computing CC-SHAP. Without this information, the CC-SHAP scores in Table 4 cannot be reproduced, and the high scores for Llama models could partly reflect lexical overlap between the explanation and the email text rather than internal decision alignment. Please provide the full prompt, generation settings, and the exact alignment method used when computing the explanation SHAP vector.
minor comments (6)
  1. [Abstract and Section 6.2] The term 'ConsistenCy' is capitalized inconsistently; consider using 'consistency' or 'CC-SHAP' uniformly.
  2. [Section 6.2] The sentence 'potentially introducing the false negatives (False Alarm) in phishing classification' uses 'false alarm' incorrectly; false negatives are missed detections, not false alarms. Please correct the terminology.
  3. [Equation (1)] In equation (1), the probability notation P is not defined; specify that it is the model's output probability for the target class (phishing or legitimate), and define the token-masking procedure more precisely.
  4. [Table 4] The accuracy columns in Table 4 are labeled '40.0', '30.0', '80.0', etc., without percentage signs; add '%' or clarify in the caption that the values are percentages.
  5. [References] Some references are incomplete or informal: reference [22] contains a typo ('Mater . Contin'), and references [23]–[25] point to Hugging Face URLs without paper titles for WizardLM-2-7B. Please provide complete bibliographic entries.
  6. [Figure 1] Figure 1 is described as a 'snapshot' but the text does not explain how SHAP values are visualized or how the figure was generated; please add a detailed caption and clarify the visual encoding.

Circularity Check

1 steps flagged · score 1.0 of 10

No significant circularity; the CC-SHAP/accuracy trade-off is measured, not derived. One mild tautology (CC-SHAP defined as alignment); the flagged 'baseline models' and same-data issues are correctness risks, not circularity.

  1. self definitional [Section 5, Eq. (3); Section 6.2 (Table 4); Abstract]
    "CC-SHAP = 1− cosine-dist(ϕ(p) norm, ϕ(e) norm) (3) ... Llama 7B and Llama 8B model exhibited high CC-SHAP scores suggesting stronger self-consistency in their prediction-explanation token alignment ... Llama models exhibit stronger prediction–explanation token alignment with higher CC SHAP scores."

    By Eq. (3), CC-SHAP is defined as 1 minus the cosine distance between the normalized prediction and explanation SHAP vectors; 'prediction-explanation token alignment' is precisely this quantity. Therefore the statement that high CC-SHAP scores indicate stronger self-consistency/token alignment restates the metric's definition rather than establishing an independent fact. The non-tautological content of the central claim is the cross-model comparison and the correlation with Table 4 accuracy, which are measured rather than forced by construction, so this step is mild and not load-bearing for the accuracy trade-off.

full rationale

The central claim — Llama models show high CC-SHAP but low phishing accuracy, Wizard shows low CC-SHAP but higher accuracy — is an empirical comparison of measured quantities (Table 4) and does not reduce to a fitted parameter renamed as a prediction, nor does it rest on a self-citation chain: [28] (Parcalabescu & Frank, the CC-SHAP source) is external, and the few co-author citations ([1], [11]) are background, not load-bearing. The only definitional step is mild: Eq. (3) defines CC-SHAP as the cosine agreement between prediction and explanation SHAP vectors, so asserting that high CC-SHAP means stronger prediction-explanation token alignment is a tautological restatement; the empirical content lies in the cross-model difference and the accuracy trade-off, which are measured, not constructed. Two flagged passages are serious but non-circular correctness risks. First, Section 6.2 states Table 4 was obtained 'using the baseline models'; if literal, the abstract's and Section 7's attribution of the trade-off to the fine-tuned models is unsupported — Table 2 shows fine-tuned Llama 7B/8B at 90.9/93.3% validation accuracy versus 40/30% in Table 4, and no fine-tuning variant is identified for Table 4. Second, Section 2 says 'a subset of the same data was also utilized to evaluate the CC-SHAP method'; with no sampling or holdout procedure given in Section 6.2, Table 4's accuracies may be in-sample, limiting generalization. Additionally, Section 7's claim that contrastive learning 'delivers no accuracy gain' is unsupported by Table 3, which reports losses only. None of these is a definitional reduction or self-citation-forced inference, so the circularity score stays at 1.

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

The empirical comparison relies on manually selected fine-tuning settings, an unreported SHAP sampling budget, an unspecified evaluation subset, and the assumption that CC-SHAP alignment reflects faithfulness. No new entities are introduced.

free parameters (3)
  • Fine-tuning hyperparameters (learning rate, epochs, LoRA rank, batch size)
    Chosen by hand but never reported; every training result in Tables 2 and 3 depends on these choices.
  • Monte Carlo sample count N for SHAP approximation
    Eq. 1 requires a sample count N per token; without it, the computed SHAP vectors and all CC-SHAP scores cannot be reproduced.
  • Evaluation subset composition = 20 phishing and 20 ham emails
    The 40 emails for the headline comparison are selected without a stated randomization or holdout procedure.
assumptions (4)
  • standard math Monte Carlo masking in Eq. 1 approximates true Shapley values for token contributions
    The paper invokes a sampling-based Shapley estimator without convergence checks or sample counts; attribution accuracy is assumed. Section 5, Eq. 1.
  • domain assumption CC-SHAP alignment between prediction and explanation SHAP vectors is a measure of faithfulness and self-consistency
    Adopted from Parcalabescu and Frank [28] and extended to phishing without validating the measure in this domain or against human judgments. Section 5.
  • ad hoc to paper Explanations generated by the fine-tuned LLM can be treated as a faithful account of its decisions
    The method computes SHAP values over explanation text but never describes how explanations are elicited or why model-generated text reveals internal reasoning. Section 5 and Figure 1.
  • ad hoc to paper The evaluation subset from Nazario and Enron is not leaked from fine-tuning
    Section 2 says a subset of the same data was used for CC-SHAP evaluation, and Section 6.2 gives no held-out split; the non-leakage assumption is unverified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evaluating Large Language Models for Phishing Detection, Self-Consistency, Faithfulness, and Explainability." pith.science (2026). https://pith.science/paper/AHUISNWA

@misc{pith2026250613746,
  author       = {Pith},
  title        = {Pith review of: Evaluating Large Language Models for Phishing Detection, Self-Consistency, Faithfulness, and Explainability},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AHUISNWA}},
  note         = {Machine review of arXiv:2506.13746}
}
read the original abstract

Phishing attacks remain one of the most prevalent and persistent cybersecurity threat with attackers continuously evolving and intensifying tactics to evade the general detection system. Despite significant advances in artificial intelligence and machine learning, faithfully reproducing the interpretable reasoning with classification and explainability that underpin phishing judgments remains challenging. Due to recent advancement in Natural Language Processing, Large Language Models (LLMs) show a promising direction and potential for improving domain specific phishing classification tasks. However, enhancing the reliability and robustness of classification models requires not only accurate predictions from LLMs but also consistent and trustworthy explanations aligning with those predictions. Therefore, a key question remains: can LLMs not only classify phishing emails accurately but also generate explanations that are reliably aligned with their predictions and internally self-consistent? To answer these questions, we have fine-tuned transformer based models, including BERT, Llama models, and Wizard, to improve domain relevance and make them more tailored to phishing specific distinctions, using Binary Sequence Classification, Contrastive Learning (CL) and Direct Preference Optimization (DPO). To that end, we examined their performance in phishing classification and explainability by applying the ConsistenCy measure based on SHAPley values (CC SHAP), which measures prediction explanation token alignment to test the model's internal faithfulness and consistency and uncover the rationale behind its predictions and reasoning. Overall, our findings show that Llama models exhibit stronger prediction explanation token alignment with higher CC SHAP scores despite lacking reliable decision making accuracy, whereas Wizard achieves better prediction accuracy but lower CC SHAP scores.

Figures

Figures reproduced from arXiv: 2506.13746 by the authors.

Figure 1
Figure 1. Snapshot of Model Input, Generated Explanation, and CC-SHAP Scores, with SHAP Values for Top Contributing Tokens During Prediction [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 22 canonical work pages

  1. [1]

    Training to detect phishing emails: Effects of the frequency of experienced phishing emails,

    K. Singh, P. Aggarwal, P. Rajivan, and C. Gonzalez, “Training to detect phishing emails: Effects of the frequency of experienced phishing emails,” in Proceedings of the human factors and er- gonomics society annual meeting , vol. 63, no. 1. SAGE Publica- tions Sage CA: Los Angeles, CA, 2019, pp. 453–457

    work page 2019

  2. [2]

    A systematic literature review on phishing email detection using natural language processing techniques,

    S. Salloum, T. Gaber, S. Vadera, and K. Shaalan, “A systematic literature review on phishing email detection using natural language processing techniques,” IEEE Access , vol. 10, pp. 65 703–65 727, 2022

    work page 2022

  3. [3]

    Why phishing works,

    R. Dhamija, J. D. Tygar, and M. Hearst, “Why phishing works,” in Proceedings of the SIGCHI conference on Human Factors in computing systems , 2006, pp. 581–590

    work page 2006

  4. [4]

    Deep learning for phishing detection: Taxonomy, current challenges and future directions,

    N. Q. Do, A. Selamat, O. Krejcar, E. Herrera-Viedma, and H. Fu- jita, “Deep learning for phishing detection: Taxonomy, current challenges and future directions,” Ieee Access, vol. 10, pp. 36 429– 36 463, 2022

    work page 2022

  5. [5]

    Chatspamdetec- tor: Leveraging large language models for effective phishing email detection,

    T. Koide, N. Fukushi, H. Nakano, and D. Chiba, “Chatspamdetec- tor: Leveraging large language models for effective phishing email detection,” arXiv preprint arXiv:2402.18093 , 2024

    arXiv 2024

  6. [6]

    Devising and detecting phishing emails using large language models,

    F. Heiding, B. Schneier, A. Vishwanath, J. Bernstein, and P. S. Park, “Devising and detecting phishing emails using large language models,” IEEE Access , 2024

    work page 2024

  7. [7]

    Enhancing Phishing Email Identification with Large Language Models

    C. Lee, “Enhancing phishing email identification with large lan- guage models,” arXiv preprint arXiv:2502.04759 , 2025

    work page Pith review arXiv 2025

  8. [8]

    An improved transformer- based model for detecting phishing, spam and ham emails: A large language model approach,

    S. Jamal, H. Wimmer, and I. H. Sarker, “An improved transformer- based model for detecting phishing, spam and ham emails: A large language model approach,” Security and Privacy , vol. 7, no. 5, p. e402, 2024

    work page 2024

Show all 29 references
  1. [9]

    Spear phishing with large language models,

    J. Hazell, “Spear phishing with large language models,” arXiv preprint arXiv:2305.06972, 2023

    2023 arXiv

  2. [10]

    Next- generation phishing: How llm agents empower cyber attackers,

    K. Afane, W. Wei, Y . Mao, J. Farooq, and J. Chen, “Next- generation phishing: How llm agents empower cyber attackers,” in 2024 IEEE International Conference on Big Data (BigData) . IEEE, 2024, pp. 2558–2567

    2024

  3. [11]

    How well does gpt phish people? an investigation involving cognitive biases and feedback,

    M. Sharma, K. Singh, P. Aggarwal, and V . Dutt, “How well does gpt phish people? an investigation involving cognitive biases and feedback,” in 2023 IEEE European Symposium on Security and Privacy Workshops (EuroS&PW) . IEEE, 2023, pp. 451–457

    2023

  4. [12]

    Expli- cate: Enhancing phishing detection through explainable ai and llm- powered interpretability,

    B. Lim, R. Huerta, A. Sotelo, A. Quintela, and P. Kumar, “Expli- cate: Enhancing phishing detection through explainable ai and llm- powered interpretability,” arXiv preprint arXiv:2503.20796 , 2025

    2025 arXiv

  5. [13]

    An explainable transformer-based model for phishing email detection: A large language model ap- proach,

    M. A. Uddin and I. H. Sarker, “An explainable transformer-based model for phishing email detection: A large language model ap- proach,” arXiv preprint arXiv:2402.13871 , 2024

    2024 arXiv

  6. [14]

    A compari- son of natural language processing and machine learning methods for phishing email detection,

    P. Bountakas, K. Koutroumpouchos, and C. Xenakis, “A compari- son of natural language processing and machine learning methods for phishing email detection,” in Proceedings of the 16th Interna- tional Conference on Availability, Reliability and Security , 2021, pp. 1–12

    2021

  7. [15]

    Phishing email corpus,

    J. Nazario, “Phishing email corpus,” 2021, accessed: April 2021. [Online]. Available: https://monkey.org/∼jose/phishing/

    2021

  8. [16]

    Enhancing phishing email detection with stylometric features and classifier stacking,

    I. Chanis and A. Arampatzis, “Enhancing phishing email detection with stylometric features and classifier stacking,” International Journal of Information Security , vol. 24, no. 1, pp. 1–16, 2025

    2025

  9. [17]

    Hackmentor: Fine-tuning large language models for cybersecurity,

    J. Zhang, H. Wen, L. Deng, M. Xin, Z. Li, L. Li, H. Zhu, and L. Sun, “Hackmentor: Fine-tuning large language models for cybersecurity,” in 2023 IEEE 22nd International Conference on Trust, Security and Privacy in Computing and Communications (TrustCom). IEEE, 2023, pp. 452–461

    2023

  10. [18]

    Data-efficient fine-tuning for llm-based recommendation,

    X. Lin, W. Wang, Y . Li, S. Yang, F. Feng, Y . Wei, and T.-S. Chua, “Data-efficient fine-tuning for llm-based recommendation,” in Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2024, pp. 365–374

    2024

  11. [19]

    Parameter- efficient fine-tuning for large models: A comprehensive survey,

    Z. Han, C. Gao, J. Liu, J. Zhang, and S. Q. Zhang, “Parameter- efficient fine-tuning for large models: A comprehensive survey,” arXiv preprint arXiv:2403.14608 , 2024

    2024 arXiv

  12. [20]

    Lora: Low-rank adaptation of large language mod- els,

    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 mod- els,” arXiv preprint arXiv:2106.09685 , 2021

    2021 arXiv

  13. [21]

    Bert: Pre- training of deep bidirectional transformers for language under- standing,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre- training of deep bidirectional transformers for language under- standing,” in Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technol...

    2019

  14. [22]

    Comparative analysis of machine learn- ing algorithms for email phishing detection using tf-idf, word2vec, and bert,

    A. Al Tawil, L. Almazaydeh, D. Qawasmeh, B. Qawasmeh, M. Al- shinwan, and K. Elleithy, “Comparative analysis of machine learn- ing algorithms for email phishing detection using tf-idf, word2vec, and bert,” Comput. Mater . Contin, vol. 81, p. 3395, 2024

    2024

  15. [23]

    Llama: Large language model meta ai,

    M. AI, “Llama: Large language model meta ai,” 2023, available on Hugging Face: https://huggingface.co/meta-llama/Llama-7b-hf

    2023

  16. [24]

    Llama 3 8b model,

    ——, “Llama 3 8b model,” https://huggingface.co/meta- llama/llama-3-8b, 2023, available on Hugging Face

    2023

  17. [25]

    Wizard 7b: dreamgen/wizardlm-2-7b,

    Dreamgen, “Wizard 7b: dreamgen/wizardlm-2-7b,” Hugging Face Repository, 2024. [Online]. Available: https://huggingface.co/ dreamgen/WizardLM-2-7B

    2024

  18. [26]

    Direct preference optimization: Your language model is secretly a reward model,

    R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn, “Direct preference optimization: Your language model is secretly a reward model,” Advances in Neural Information Processing Systems, vol. 36, pp. 53 728–53 741, 2023

    2023

  19. [27]

    Contrastive repre- sentation learning: A framework and review,

    P. H. Le-Khac, G. Healy, and A. F. Smeaton, “Contrastive repre- sentation learning: A framework and review,” Ieee Access , vol. 8, pp. 193 907–193 934, 2020

    2020

  20. [28]

    On measuring faithfulness or self-consistency of natural language explanations,

    L. Parcalabescu and A. Frank, “On measuring faithfulness or self-consistency of natural language explanations,” arXiv preprint arXiv:2311.07466, 2023

    2023 arXiv

  21. [29]

    Can machines think like humans? a behavioral evaluation of llm-agents in dictator games,

    J. Ma, “Can machines think like humans? a behavioral evaluation of llm-agents in dictator games,” arXiv preprint arXiv:2410.21359, 2024

    2024

Pith tools

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