Pith. sign in

REVIEW 3 major objections 5 minor 34 references

No for Some, Yes for Others: Persona Prompts and Other Sources of False Refusal in Language Models

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Persona prompts change false refusal rates less than model and task choice do.

desk verdict Solid multi-factor empirical ranking of false-refusal sources (model > task > persona > prompt), with an untested refusal detector and missing artifacts keeping it short of a reference result. read the letter →

arxiv 2509.08075 v1 pith:ZHE74RHF submitted 2025-09-09 cs.CL

Flor Miriam Plaza-del-Arco , Paul Röttger , Nino Scherrer , Emanuele Borgonovo , Elmar Plischke , Dirk Hovy This is my paper
classification cs.CL
keywords falserefusalpersonapromptingsociodemographicpersonassafetyalignmentMonteCarlosensitivityanalysisWassersteindistanceover-refusalLLMevaluation
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 tries to establish that sociodemographic persona prompting is a real but secondary cause of false refusals in language models. Across 16 open-weight models, three classification tasks, and nine prompt paraphrases, the model itself and the task explain far more variation in refusal rates than the persona does. The authors argue that earlier studies overstated persona effects because they did not control for task, model generation, and prompt phrasing. If correct, the practical lesson is that refusal behavior should be attributed mainly to model safety alignment and task sensitivity, not to the user's stated identity.

What carries the argument

The central mechanism is a nested Monte Carlo sampling scheme that randomly pairs each input text with a persona and a prompt template, so that the combinatorial space of persona×prompt×model×task is explored in a balanced, sample-efficient way. Quantification then rests on a squared Wasserstein-2 global sensitivity index, computed via optimal transport, which measures how much knowing each factor shifts the refusal distribution; a regularized logistic regression on refusal outcome confirms the ranking and gives per-level coefficients.

What would settle it

Hand-label a random sample of model responses across all personas, tasks, and models, and compare human refusal judgments with the string-match detector; if disagreements are concentrated in particular personas or models, the reported refusal rates and the model-vs-persona ranking would shift.

Watch

Extended reading notes

Core claim

The paper claims that false refusals in LLM classification tasks are driven primarily by model choice, secondarily by task, third by sociodemographic persona, and least by prompt paraphrase. Using a Monte Carlo sampling design and a Wasserstein-distance sensitivity analysis, the authors rank these four sources of refusal variation. They also find a generational pattern: newer Llama and Qwen models refuse far less, while the newest Gemma2 models refuse more, and offensiveness classification triggers far more refusals than NLI. Certain personas—Black, White, transgender, and Muslim—elicit disproportionately more refusals in some models, which the paper reads as evidence of bias in safety align

Load-bearing premise

The automatic refusal detector only counts responses that begin with a fixed list of refusal phrases, so refusals phrased differently would be missed; the paper acknowledges this and assumes the misses are spread evenly enough not to change the factor ranking.

Editorial extensions

If this is right

  • Future studies of persona effects must control for model, task, and prompt design, or they risk attributing to personas what is actually caused by other experimental choices.
  • Refusal rates in classification tasks decrease substantially in newer model generations within some families, but not all, so progress on false refusal is not uniform.
  • Tasks involving sensitive content, especially offensiveness classification, will keep eliciting far more false refusals than objective tasks like NLI.
  • Open-ended or unforced prompts increase refusals; forcing a label output suppresses most refusals, which is an experimental-design choice that can change reported results dramatically.
  • Some sociodemographic groups experience systematically higher false refusals in particular models, suggesting that safety tuning has fairness implications that should be audited per model rather than assumed from persona labels.

Reading between the lines

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

  • An interaction between persona and model generation may be the real story: the same persona label moves refusal rates strongly in older models and barely at all in newer ones, so a single 'persona effect' number will be misleading without specifying the model generation.
  • The string-match refusal detector could be tested directly against human annotation; if modern models refuse in longer, hedged forms that do not start with the listed phrases, the reported rates and the ranking of factors would need revision, although the paper argues any such miss is likely evenly distributed.
  • The same Monte Carlo plus Wasserstein sensitivity pipeline could be transferred to other outcome variables, such as answer accuracy, sycophancy, or formatting compliance, to separate which experimental controls actually drive behavior.
  • The logistic regression's persona coefficients are main effects only; whether Black, Muslim, and transgender personas remain the top-refusal groups in multimodal, larger, or proprietary models is an open empirical question.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper reports a large-scale controlled study of false refusals in open-weight LLMs. The authors prompt 16 models with 15 sociodemographic personas across three classification tasks (NLI, politeness, offensiveness) and nine prompt paraphrases, using a Monte Carlo sampling scheme to keep the evaluation feasible. They quantify the contribution of model, task, persona, and prompt to false refusal rates using a Wasserstein-distance sensitivity index and a logistic regression. The central finding is that model choice is the most influential factor, followed by task, then persona, then prompt; the authors conclude that persona effects on false refusal have been overestimated in prior work, while model safety behavior and sensitive-task content dominate. They also report that newer Llama/Qwen models refuse less, while Gemma2 models show increased refusal rates, and that certain personas (Black, White, transgender, Muslim) are associated with more refusals in some models.

Significance. If the results hold, the paper provides a valuable corrective to the existing persona-refusal literature: it controls for several confounds simultaneously and quantifies their relative importance rather than reporting pairwise persona differences. The Monte Carlo sampling design and the use of a global sensitivity measure are appropriate for the combinatorial search space, and the inclusion of 16 models and three tasks is a strength. The main quantitative conclusion—model and task dominate persona—is plausible and would be practically useful for researchers designing fairness or safety evaluations. However, the paper's central claim is only as strong as its refusal detector, and the detector is neither validated nor released; this, together with the lack of code/data and uncertainty quantification, currently limits the confidence that can be placed in the ranking. The paper also makes a testable claim about the distribution of detector misses, but does not test it.

major comments (3)
  1. [§3.3 and Appendix C] The refusal detector is load-bearing for the paper's central claim, and it has both false-negative and false-positive risks. It flags only responses beginning with a fixed list (e.g., 'sorry', 'i cannot', 'as an ai') as refusals; any refusal phrased differently (e.g., 'I'd rather not', 'This cannot be classified') is missed. Conversely, the list includes 'no,' and 'no.' as refusal markers, so a legitimate negative answer such as 'No, a man would not label this as offensive' would be counted as a refusal. The Limitations section concedes that refusal rates 'might thus be marginally higher than reported, but likely to be evenly enough distributed to not change results,' but this is asserted, not demonstrated. Since output format and phrasing interact with prompt strictness (unforced vs. semi-forced/forced) and with model generation style, differential misclassification across models or per
  2. [§3.1–3.2] The Monte Carlo sampling procedure is described only at a high level. The authors say they 'sample at random' tuples of (prompt template, persona) and 'ensure a balanced distribution,' but they do not provide the exact algorithm, the number of samples drawn per combination, the random seed, or any pseudocode. Without this information, and without code or data release, the reported refusal rates and sensitivity indices cannot be reproduced. The paper's reproducibility claim would be substantially strengthened by releasing the sampling code, the generated prompts, and the model outputs (or an anonymized subset), or at least by specifying the full procedure in an appendix.
  3. [§4.1, Table 1 and §4.6.1, Figure 5] The refusal rates and the Wasserstein importance indices are reported as point estimates without confidence intervals. The paper's ranking (model > task > persona > prompt) is a comparative claim, and it is not shown that the differences between adjacent factors, especially persona vs. prompt, are statistically distinguishable. The authors should provide bootstrap or exact binomial confidence intervals for the refusal rates in Table 1 and for the importance indices in Figure 5. This is particularly relevant for near-zero rates (e.g., many models in the NLI condition), where small absolute differences can be driven by a handful of responses.
minor comments (5)
  1. [§4.4] The text states that false refusal rates are 'much higher in the persona setup (14.68%),' but the persona-free average is not reported in the text. Please give both numbers and the delta explicitly, or refer to the figure with the actual values.
  2. [Table 7] The prompt coefficient 'pp5' appears to be a typo for 'p5'; the prompt identifiers are otherwise p1–p9. Please align the labels between the table and the appendix prompt templates.
  3. [Appendix D.2] The caption says 'Figures 7, 8 show the variation...' but it is referring to Figures 10 and 11. The cross-reference is wrong.
  4. [Author affiliations] The affiliation for Emanuele Borgonovo is spelled 'Boconni University' in the author block; it should be 'Bocconi University'.
  5. [Figure 6 / §4.6.2] The text says 'Llama3.8B' when referring to Llama3-8B. Please fix the model name for consistency with Table 1.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the refusal-rate measurements and importance rankings are empirical outputs of a stated detector and statistical tools, not re-statements of their inputs.

full rationale

The paper is an empirical measurement study. Refusal is operationalized by the explicit prefix-string detector in Appendix C; the Wasserstein and logistic-regression analyses in §4.6 are applied to the resulting binary labels. There is no equation in which the reported importance ranking is defined in terms of itself: the detector's prefix list is fixed before the experiments and does not encode model/task/persona/prompt as factors. The sensitivity measure from Borgonovo et al. is cited from prior work by co-authors, but it is a generic statistical tool with stated formulas, not a uniqueness theorem or an ansatz that pre-selects the conclusion. The prompt typology from Röttger et al. (2024a) is an experimental design choice, and the finding that unforced prompts elicit more refusals is an empirical output, not an assumption. The Limitations section's caveat that the detector 'might not have identified the full range of refusals' is a measurement-validity concern; it does not make the ranking circular, because the analysis would be equally well-defined for any detector and the conclusion is not logically entailed by the detector's construction. No fitted parameter is relabeled as a prediction. Overall, the central claims are self-contained relative to their operational definitions.

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

The paper introduces no new theoretical quantities or entities. Its claims rest on standard statistical assumptions and on design choices (sample sizes, persona set, prompt set, refusal detector) that are only partially justified.

free parameters (3)
  • Number of persona/prompt variations per input sample = 10
    Chosen by hand; every selected input is expanded into 10 variations. This sample size is not justified and could affect the precision of the importance estimates.
  • Number of samples per class in each task = 1000
    Selected by hand. The paper does not report a power analysis or confidence intervals around the refusal rates, so the adequacy of this sample size is unverified.
  • Logistic regression regularization strength
    The paper says 'regularized logistic regression' but does not report the regularization type, strength, or fitting details. This could affect the extreme coefficients reported.
assumptions (4)
  • domain assumption Refusals are detectable by a fixed list of starting phrases such as "I'm sorry" or "I cannot".
    Used in Section 3.3 and Appendix C for all refusal classification. The paper itself acknowledges in Limitations that this may miss some refusals.
  • domain assumption The three tasks (NLI, politeness, offensiveness) are representative of safe classification tasks where false refusals are expected to be low.
    The choice of tasks in Section 2.2 drives the entire refusal rate distribution; if a different task set were chosen, the relative importance ranking could differ.
  • domain assumption Temperature-0 greedy decoding is deterministic and adequate for measuring refusal behavior.
    They set temperature to 0 in Section 3.2, but do not test sensitivity to sampling temperature, which they list as an untested factor in Limitations.
  • domain assumption The Monte Carlo sampling scheme yields an unbiased, balanced estimate of factor importance.
    The construction of balanced samples in Section 3.1 is described only briefly; no proof or diagnostic is given that the 10-variation expansion produces unbiased estimates over the full combinatorial space.

how reviews work

0 comments
Cite this review

Pith. "Pith review of No for Some, Yes for Others: Persona Prompts and Other Sources of False Refusal in Language Models." pith.science (2026). https://pith.science/paper/ZHE74RHF

@misc{pith2026250908075,
  author       = {Pith},
  title        = {Pith review of: No for Some, Yes for Others: Persona Prompts and Other Sources of False Refusal in Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZHE74RHF}},
  note         = {Machine review of arXiv:2509.08075}
}
read the original abstract

Large language models (LLMs) are increasingly integrated into our daily lives and personalized. However, LLM personalization might also increase unintended side effects. Recent work suggests that persona prompting can lead models to falsely refuse user requests. However, no work has fully quantified the extent of this issue. To address this gap, we measure the impact of 15 sociodemographic personas (based on gender, race, religion, and disability) on false refusal. To control for other factors, we also test 16 different models, 3 tasks (Natural Language Inference, politeness, and offensiveness classification), and nine prompt paraphrases. We propose a Monte Carlo-based method to quantify this issue in a sample-efficient manner. Our results show that as models become more capable, personas impact the refusal rate less and less. Certain sociodemographic personas increase false refusal in some models, which suggests underlying biases in the alignment strategies or safety mechanisms. However, we find that the model choice and task significantly influence false refusals, especially in sensitive content tasks. Our findings suggest that persona effects have been overestimated, and might be due to other factors.

Figures

Figures reproduced from arXiv: 2509.08075 by the authors.

Figure 1
Figure 1. Comparison of refusal rates (%) by model in [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Variation of refusal rates (%) per [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Refusal rates (%) of models across the 15 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: Variable Importance through Wasserstein Dis [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Top 10 positive and negative regression coeffi [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Refusal rates (%) for the NLI task across personas. physically-disabled able-bodied black white asian man woman non-binary transgender man transgender woman atheist christian jewish muslim religious Llama2-7B Llama2-13B Llama3-8B Llama3.1-8B Llama3.2-1B Llama3.2-3B Qwe…
Figure 8
Figure 8. Figure 8: Refusal rates (%) for the politeness task across personas. physically-disabled able-bodied black white asian man woman non-binary transgender man transgender woman atheist christian jewish muslim religious Llama2-7B Llama2-13B Llama3-8B Llama3.1-8B Llama3.2-1B Llama3.2…
Figure 9
Figure 9. Figure 9: Refusal rates (%) for the offensiveness task across personas. D.2 Refusal by Prompt Figures 7, 8 show the variation of false refusals across models and prompt strictness response lev- [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 11
Figure 11. Figure 11: Refusal rates (%) across models for the po [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 10
Figure 10. Figure 10: Refusal rates (%) across models for the NLI [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 14 canonical work pages

  1. [1]

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

    2024

  2. [2]

    Bang An, Sicheng Zhu, Ruiyi Zhang, Michael-Andrei Panaitescu-Liess, Yuancheng Xu, and Furong Huang. 2024. Automatic pseudo-harmful prompt generation for evaluating false refusals in large language models. In ICML 2024 Next Generation of AI Safety Workshop

    work page 2024

  3. [3]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, and 1 others. 2023. Qwen technical report. arXiv preprint arXiv:2309.16609

    arXiv 2023

  4. [4]

    Emanuele Borgonovo, Alessio Figalli, Elmar Plischke, and Giuseppe Savar \'e . 2024. Global sensitivity analysis via optimal transport. Management Science. Online First

    work page 2024

  5. [5]

    Emanuele Borgonovo, Valentina Ghidini, Roman Hahn, and Elmar Plischke. 2023. Classifier explainability with measures of statistical association. Computational Statistics and Data Analysis, 182:197701/1--16

    work page 2023

  6. [6]

    Hazen, and Elmar Plischke

    Emanuele Borgonovo, Gordon B. Hazen, and Elmar Plischke. 2016. A common rationale for global sensitivity measures and their estimation. Risk Analysis, 36(10):1871--1895

    work page 2016

  7. [7]

    Khaoula Chehbouni, Megha Roshan, Emmanuel Ma, Futian Wei, Afaf Taik, Jackie Cheung, and Golnoosh Farnadi. 2024. https://doi.org/10.18653/v1/2024.findings-acl.927 From representational harms to quality-of-service harms: A case study on llama 2 safety safeguards . In Findings of the Association for Computational Linguistics: ACL 2024, pages 15694--15710, Ba...

    work page doi:10.18653/v1/2024.findings-acl.927 2024

  8. [8]

    Georgiou, and Michele Pavon

    Yongxin Chen, Tryphon T. Georgiou, and Michele Pavon. 2021. Stochastic control liaisons: R ichard S inkhorn meets G aspard M onge on a S chr \"o dinger bridge. SIAM Review, 63(2):249--313

    work page 2021

Show all 34 references
  1. [9]

    Alexis Conneau, Ruty Rinott, Guillaume Lample, Adina Williams, Samuel Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. https://doi.org/10.18653/v1/D18-1269 XNLI : Evaluating cross-lingual sentence representations . In Proceedings of the 2018 Conference on Empirical Methods ...

    2018 doi

  2. [10]

    Justin Cui, Wei-Lin Chiang, Ion Stoica, and Cho-Jui Hsieh. 2024. Or-bench: An over-refusal benchmark for large language models. arXiv preprint arXiv:2405.20947

    2024 arXiv

  3. [11]

    Pedro Henrique Luz de Araujo and Benjamin Roth. 2024. Helpful assistant or fruitful facilitator? investigating how personas affect language model behavior. arXiv preprint arXiv:2407.02099

    2024 arXiv

  4. [12]

    Prannaya Gupta, Le Qi Yau, Hao Han Low, I-Shiang Lee, Hugo Maximus Lim, Yu Xin Teoh, Koh Jia Hng, Dar Win Liew, Rishabh Bhardwaj, Rajat Bhardwaj, and Soujanya Poria. 2024 a . https://doi.org/10.18653/v1/2024.emnlp-demo.42 W alled E val: A comprehensive safety evaluation toolki...

    2024 doi

  5. [13]

    Shashank Gupta, Vaishnavi Shrivastava, Ameet Deshpande, Ashwin Kalyan, Peter Clark, Ashish Sabharwal, and Tushar Khot. 2024 b . Bias R uns D eep: Implicit reasoning biases in persona-assigned LLM s. In The Twelfth International Conference on Learning Representations

    2024

  6. [14]

    Guangyuan Jiang, Manjie Xu, Song-Chun Zhu, Wenjuan Han, Chi Zhang, and Yixin Zhu. 2023. Evaluating and inducing personality in pre-trained language models. Advances in Neural Information Processing Systems, 36:10622--10643

    2023

  7. [15]

    Meta. 2024. https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/ Llama 3.2: Revolutionizing edge ai and vision with open, customizable models

    2024

  8. [16]

    Jiaxin Pei and David Jurgens. 2023. https://doi.org/10.18653/v1/2023.law-1.25 When do annotator demographics matter? measuring the influence of annotator demographics with the POPQUORN dataset . In Proceedings of the 17th Linguistic Annotation Workshop (LAW-XVII), pages 252--2...

    2023 doi

  9. [17]

    Flor Miriam Plaza-del-Arco , Amanda Cercas Curry, Susanna Paoli, Alba Cercas Curry, and Dirk Hovy. 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.251 Divine LL a MA s: Bias, stereotypes, stigmatization, and emotion representation of religion in large language models . I...

    2024 doi

  10. [18]

    Omid Rafieian and Hema Yoganarasimhan. 2023. Ai and personalization. Artificial Intelligence in Marketing, pages 77--102

    2023

  11. [19]

    Paul R \"o ttger, Valentin Hofmann, Valentina Pyatkin, Musashi Hinck, Hannah Kirk, Hinrich Schuetze, and Dirk Hovy. 2024 a . https://doi.org/10.18653/v1/2024.acl-long.816 Political compass or spinning arrow? towards more meaningful evaluations for values and opinions in large ...

    2024 doi

  12. [20]

    Paul R \"o ttger, Hannah Kirk, Bertie Vidgen, Giuseppe Attanasio, Federico Bianchi, and Dirk Hovy. 2024 b . https://doi.org/10.18653/v1/2024.naacl-long.301 XST est: A test suite for identifying exaggerated safety behaviours in large language models . In Proceedings of the 2024...

    2024 doi

  13. [21]

    Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. 2024. https://doi.org/10.18653/v1/2024.acl-long.399 L a MP : When large language models meet personalization . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volu...

    2024 doi

  14. [22]

    Nino Scherrer, Claudia Shi, Amir Feder, and David Blei. 2023. Evaluating the moral beliefs encoded in llms. Advances in Neural Information Processing Systems, 36:51778--51809

    2023

  15. [23]

    Melanie Sclar, Yejin Choi, Yulia Tsvetkov, and Alane Suhr. 2023. Quantifying language models' sensitivity to spurious features in prompt design or: How i learned to start worrying about prompt formatting. arXiv preprint arXiv:2310.11324

    2023 arXiv

  16. [24]

    Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivi \`e re, Mihir Sanjay Kale, Juliette Love, and 1 others. 2024 a . Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295

    2024 arXiv

  17. [25]

    Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, L \'e onard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ram \'e , and 1 others. 2024 b . Gemma 2: Improving open language models at a practical size. arXiv preprint arX...

    2024 arXiv

  18. [26]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, and 1 others. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288

    2023 arXiv

  19. [27]

    C \'e dric Villani. 2009. Optimal Transport: Old and New. Springer Verlag, Berlin

    2009

  20. [28]

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, and 1 others. 2024 a . Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191

    2024 arXiv

  21. [29]

    Xinpeng Wang, Chengzhi Hu, Paul R \"o ttger, and Barbara Plank. 2024 b . Surgical, cheap, and flexible: Mitigating false refusal in language models via single vector ablation. arXiv preprint arXiv:2410.03415

    2024 arXiv

  22. [30]

    Johannes C. W. Wiesel. 2022. Measuring association with W asserstein distances. Bernoulli, 28(4):2816--2832

    2022

  23. [31]

    Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. https://doi.org/10.18653/v1/N18-1101 A broad-coverage challenge corpus for sentence understanding through inference . In Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computation...

    2018 doi

  24. [32]

    Zhehao Zhang, Ryan A Rossi, Branislav Kveton, Yijia Shao, Diyi Yang, Hamed Zamani, Franck Dernoncourt, Joe Barrow, Tong Yu, Sungchul Kim, and 1 others. 2024. Personalization of large language models: A survey. arXiv preprint arXiv:2411.00027

    2024 arXiv

  25. [33]

    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...

  26. [34]

    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 4, 2026 · model on record in the stance chip above.