Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

Not All Jokes Land: Evaluating Large Language Models Understanding of Workplace Humor

T0 review · 4 major / 5 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 Large language models often misjudge whether workplace humor is appropriate, and the failure is sharpest for offensive and neutral jokes.

desk verdict A genuinely new workplace-humor dataset, but the evaluation prompt drops the context field the dataset is built around, so the headline result may be an artifact of under-specification. read the letter →

arxiv 2506.01819 v2 pith:UPSFSF3E submitted 2025-06-02 cs.CL cs.CY

Mohammadamin Shafiei , Hamidreza Saffari This is my paper
classification cs.CLcs.CY
keywords workplacehumorLLMevaluationappropriatenessclassificationdetectionprofessionaldatasetoffensivelanguageculturalreferencehuman-LLMalignment
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 introduces a new resource for studying humor in industrial workplaces: 304 humorous sentences, each human-annotated for appropriateness (Offensive, Mildly Inappropriate, Neutral, or Wholesome), industry context, and humor type. The authors then ask five large language models to classify each sentence's appropriateness and find that the models often fail. The lowest scores cluster in the Offensive and Neutral categories, with average weighted F1 around 0.32 and 0.31 on one generated subset, while Wholesome humor is classified far more reliably. The takeaway is that current LLMs lack the situational and cultural sensitivity needed to judge whether a workplace joke is safe, a gap that matters as AI assistants take on email and other professional writing.

What carries the argument

The load-bearing instrument is a dataset of 304 professionally situated humorous sentences, each carrying a four-level appropriateness label (Offensive, Mildly Inappropriate, Neutral, Wholesome), an industry-specific context, and a humor type. Offensive is defined as humor that targets a nation, culture, or group through biased content or harsh tone; Neutral is defined as factual, bias-free observation. These human labels serve as the ground truth against which five zero-temperature LLM classifiers are scored. Because each sentence is tied to a context and a humor type, the dataset supports fine-grained error analysis, such as the observation that cultural-reference humor is systematically missed as offensive.

What would settle it

Have a larger, gender-balanced, culturally diverse group re-annotate the 304 sentences and compute inter-annotator agreement; low agreement or substantially shifted labels would show that the Table 2 scores are not a stable measure of LLM humor judgment. A second, complementary check is to give a model a few in-context examples of clearly offensive cultural humor and see whether its Offensive F1 rises, which would indicate the paper measured a prompting limitation rather than an intrinsic deficit.

Watch

Extended reading notes

Core claim

The central claim is that LLMs often judge workplace humor appropriateness inaccurately, and the paper backs this claim with a purpose-built benchmark. On the subset of jokes generated by one LLM and labeled by human annotators, the best model reaches a weighted average F1 of 0.60 while other models range from 0.41 to 0.58; offensive humor is hardest, averaging 0.32 weighted F1 across models, and neutral humor is nearly as hard at 0.31. The pattern repeats on a second generated subset, where offensive humor averages 0.26. Wholesome humor, by contrast, is classified well, with average weighted F1 near 0.84 on the first subset. The paper attributes the failure to models overlooking cultural and national references that make humor offensive, and to the inherently broad boundaries of Neutral humor.

Load-bearing premise

The evaluation rests on the human appropriateness labels being correct and consistent; if the two annotators' judgments are noisy or skewed by their shared background, the low model scores measure disagreement with those particular labels rather than a general failure to understand workplace humor.

Editorial extensions

If this is right

  • AI assistants that draft workplace emails or posts would let offensive, culturally targeted jokes through as Neutral when left to their own judgment.
  • The benchmark provides a concrete target for alignment work: raising Offensive and Neutral F1 on this dataset would be measurable progress on humor judgment.
  • Because Wholesome humor is easy for all tested models, the practical bottleneck is detecting unsafe edge cases rather than recognizing obviously safe humor.
  • Generic toxicity filters are unlikely to close the gap, since the errors involve culturally specific references that the models currently classify as factually neutral.

Reading between the lines

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

  • Beyond the paper, the results imply that alignment training gives models little purchase on culturally encoded offense; a direct test would be to paraphrase the offensive jokes with different cultural referents and see whether model labels remain stable.
  • Beyond the paper, the annotators removed all gender-referencing sentences, so the benchmark cannot say whether the offensive-category failure extends to sexist workplace humor; an expanded, gender-balanced annotation pass would settle this.
  • Beyond the paper, the absence of inter-annotator agreement reporting means the human labels are an unexamined link in the measurement chain; computing agreement on a re-annotated subset would tell readers how much of the model error is label noise.
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 / 5 minor

Summary. The paper introduces a new dataset of 304 LLM-generated workplace-humor sentences, each annotated with one of four appropriateness levels (Offensive, Mildly Inappropriate, Neutral, Wholesome) and provided with metadata such as industry-specific context and humor type. The authors evaluate five LLMs—GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Flash, Llama-3.2-1B-Instruct, and Qwen2.5-72B-Instruct—on the task of classifying appropriateness, with the two generator models tested only on the subset they did not generate. Results in Table 2 show that Wholesome is the easiest category (average F1 around 0.84) and Offensive and Neutral are hardest (average F1 around 0.32 and 0.31 on the Claude subset), leading to the abstract claim that LLMs often struggle to judge the appropriateness of humor accurately.

Significance. If the measurement is valid, the paper provides a useful first benchmark for an underexplored aspect of human-aligned AI: the judgment of workplace humor appropriateness. Strengths include the public release of the dataset, a cross-subset evaluation design that mitigates the most direct form of circularity (the generator models are not tested on their own outputs), deterministic zero-temperature evaluation, and a per-humor-type breakdown that gives a starting point for error analysis. The paper is also appropriately cautious in its limitations section about dataset size and annotator demographics. However, the central claim depends on two load-bearing assumptions that are not currently established: that the human ground-truth labels are reliable, and that the evaluation protocol gives models the information the paper itself defines as part of the task.

major comments (4)
  1. [§4 and Appendix B] The evaluation prompt in Appendix B supplies only the sentence and the four definitions of appropriateness; it never provides the 'Industry-Specific Context' field that Section 3.1 defines as a dataset feature. This matters because the category definitions are context-sensitive: 'Mildly Inappropriate' is defined with respect to a role or department within the company, and 'Neutral' is defined as factual observation about the company. Without knowing the intended industry setting, a model cannot be expected to know, for example, whether a remark targets a department that exists in that workplace. The low Offensive and Neutral F1 scores in Table 2 may therefore reflect an under-specified classification task rather than an inability to judge workplace humor. Please rerun the evaluation with the Industry-Specific Context field included, or provide a per-item analysis showing which of the 304 items are classifiable without it, and adjust the abstract and conclusions accordingly.
  2. [§3.3] The annotation section reports that two male annotators independently assessed each sample and that a third male reviewer resolved disagreements, but it reports no inter-annotator agreement statistic, such as Cohen's kappa, for any annotation field. With only 304 items and four label categories, even moderate annotator disagreement could change the per-class F1 scores in Table 2 substantially. Please report agreement statistics separately for appropriateness, industry context, and humor type, and describe the distribution of disagreements that required the third reviewer, so that the validity of the ground-truth labels can be assessed.
  3. [§5, Table 2] The comparative claims about models are made without any measure of uncertainty. For example, the statement that Llama-3.2-1B-Instruct 'performs better than Qwen and Gemini' rests on weighted F1 scores of 0.58 versus 0.54 and 0.48 on the Claude-generated subset, computed over 156 items. With this sample size, those differences may not be statistically reliable. Please provide bootstrap confidence intervals, McNemar tests, or another appropriate significance test for the weighted F1 differences, or soften the comparative conclusions to avoid over-interpreting small differences.
  4. [§3.2 and §8] The dataset is generated entirely by LLMs, and the authors acknowledge in Section 8 that LLM-generated humor may differ from naturally occurring workplace humor. This is a limitation rather than an error, but it should be reflected more prominently in the abstract and introduction: the paper currently claims that LLMs struggle with 'workplace humor' generally, whereas the evaluated items are machine-generated workplace-humor-like sentences. An explicit statement that the dataset is a synthetic instrument, not a sample of real workplace communication, would make the scope of the claim accurate.
minor comments (5)
  1. [Abstract] The abstract ends with the orphan citation 'Etxaniz et al. (2024)' that is not connected to any sentence; remove it or integrate it into the text.
  2. [§3.1] The sentence 'with for levels' should read 'with four levels'.
  3. [Appendix A] The generation prompt says 'Professional industries-specific Humor'; this should be 'industry-specific humor' for consistency with the rest of the paper.
  4. [Table 4] The abbreviation 'Self-dep.' is not expanded in the table caption or the text; define it as 'Self-deprecation'.
  5. [§2.1 and §2.2] The related work would benefit from a comparison with any existing humor-in-context datasets beyond puns and one-liners, since the paper claims to introduce the 'first' dataset specifically for industrial settings; a brief justification of that 'first' claim is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the F1 results are genuine cross-subset measurements against human labels, not outputs forced by the dataset construction or by any self-citation chain.

full rationale

The paper's claimed chain is simple: build a 304-item, human-annotated dataset of workplace humor, then report LLM F1 scores on the appropriateness classification task; the only load-bearing evidence is Table 2. No parameter is fitted, no uniqueness theorem is invoked, and no self-citation carries weight: the reference list contains no work by Shafiei or Saffari, and the dangling citation in the abstract is to BertaQA (Etxaniz et al., 2024), an unrelated prior culture-knowledge benchmark. The one plausible circularity, evaluating models on jokes they generated themselves, is removed by construction in Section 4: 'each was only tested on the subset they were not involved in generating,' so Claude and GPT-4o scores are cross-model measurements, and the other models had no role in generation at all. The appropriateness definitions (Offensive, Mildly Inappropriate, Neutral, Wholesome) appear identically in the generation prompt (Appendix A), the annotation step (Section 3.3), and the evaluation prompt (Appendix B), but this reuse does not make any label follow by construction; models frequently disagree with the human labels, with Table 2 showing F1 values as low as 0.00 (e.g., Qwen2.5 on Offensive and Neutral in the GPT-4o subset), so the reported scores are real empirical outcomes rather than the definitions recycled. Two reviewer concerns are genuine but are not circularity: Section 3.3 reports no inter-annotator agreement statistic, so human-label reliability is unquantified, and Section 4 with Appendix B pairs each sentence with a generic prompt that omits the dataset's 'Industry-Specific Context' field, so low scores may reflect an under-specified task rather than a general inability to judge workplace humor; both are measurement-validity risks for the correctness pass, not self-referential reasoning. A bibliographic slip (Tikhonov and Shtykovskiy 2024a and 2024b citing the same paper twice) is an editing issue, not a circular step. The acknowledged limitations (dataset size, coarse humor-type grouping, all-male annotators with gender-reference removal) are stated honestly and do not hide any input-output equivalence. Because no prediction in the paper reduces to its own inputs by definition or by fit, the circularity score is 0.

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

The paper has no free parameters in the mathematical sense. It relies on several domain assumptions: the validity of the four-level appropriateness scale, the representativeness of LLM-generated humor as workplace humor, the reliability of the two-annotator labels, and the choice of weighted F1 as the evaluation metric. None of these is independently validated in the paper.

assumptions (4)
  • domain assumption The four-level appropriateness scale (Offensive, Mildly Inappropriate, Neutral, Wholesome) is an adequate and exhaustive operationalization of humor appropriateness in workplaces.
    The categories are defined by the authors in Section 3.1 and used both for annotation and evaluation; no validation against an external or theoretical standard is provided.
  • domain assumption LLM-generated humorous sentences are representative of professional industrial humor.
    All 304 sentences were produced by Claude 3.5 Sonnet or ChatGPT-4o via prompting (Section 3.2); no real-world workplace humor corpus is used, so generalizability to actual workplace banter is assumed.
  • domain assumption The two annotators' judgments, with a third reviewer for conflicts, constitute reliable ground truth.
    Section 3.3 states two male annotators assessed samples and a third reviewer resolved disagreements, but no inter-annotator agreement statistic (e.g., Cohen's kappa) is reported.
  • domain assumption Weighted average F1 over the four classes is an appropriate aggregate metric for comparing model performance.
    Used in Table 2; no justification is given for this choice over macro-F1 or accuracy, and class imbalances exist (e.g., only 21 Offensive examples in the Claude subset).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Not All Jokes Land: Evaluating Large Language Models Understanding of Workplace Humor." pith.science (2026). https://pith.science/paper/UPSFSF3E

@misc{pith2026250601819,
  author       = {Pith},
  title        = {Pith review of: Not All Jokes Land: Evaluating Large Language Models Understanding of Workplace Humor},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UPSFSF3E}},
  note         = {Machine review of arXiv:2506.01819}
}
read the original abstract

With the recent advances in Artificial Intelligence (AI) and Large Language Models (LLMs), the automation of daily tasks, like automatic writing, is getting more and more attention. Hence, efforts have focused on aligning LLMs with human values, yet humor, particularly professional industrial humor used in workplaces, has been largely neglected. To address this, we develop a dataset of professional humor statements along with features that determine the appropriateness of each statement. Our evaluation of five LLMs shows that LLMs often struggle to judge the appropriateness of humor accurately.

Figures

Figures reproduced from arXiv: 2506.01819 by the authors.

Figure 1
Figure 1. Examples of humor misclassified by LLMs. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. HumorRank: A Tournament-Based Leaderboard for Evaluating Humor Generation in Large Language Models

    cs.CL 2026-03 unverdicted novelty 7.0 of 10

    HumorRank ranks nine LLMs on textual humor using GTVH-grounded pairwise tournaments and Adaptive Swiss aggregation on the SemEval-2026 MWAHAHA dataset, finding that comedic mechanism mastery matters more than scale.

  2. HumorGen: Cognitive Synergy for Humor Generation in Large Language Models via Persona-Based Distillation

    cs.CL 2026-03 unverdicted novelty 6.0 of 10

    Persona-based Mixture-of-Thought data curation lets a 7B student outperform larger models on humor generation, while DPO and O-GRPO add no gain over SFT.

  3. One Joke to Rule them All? On the (Im)possibility of Generalizing Humor

    cs.CL 2025-08 conditional novelty 6.0 of 10

    LLMs fine-tuned on one to three humor datasets transfer partially to unseen humor types (up to 75% accuracy); diverse training helps modestly, and dad jokes enable transfer best but resist it as a target.

Reference graph

Works this paper leans on

40 extracted references · 25 canonical work pages · cited by 3 Pith papers

  1. [1]

    John A Banas, Norah Dunbar, Dariela Rodriguez, and Shr-Jie Liu. 2011. A review of humor in educational settings: Four decades of research. Communication Education, 60(1):115--144

    work page 2011

  2. [2]

    Adrian Bardon. 2005. The philosophy of humor. Comedy: A geographic and historical guide, 2:462--476

    work page 2005

  3. [3]

    Ronald A Berk. 2015. The greatest veneration: Humor as a coping strategy for the challenges of aging. Social Work in Mental Health, 13(1):30--47

    work page 2015

  4. [4]

    Kim Binsted and Graeme Ritchie. 1994. https://arxiv.org/abs/cmp-lg/9406022 An implemented model of punning riddles . Preprint, arXiv:cmp-lg/9406022

    work page Pith review arXiv 1994

  5. [5]

    Yang Chen, Chong Yang, Tu Hu, Xinhao Chen, Man Lan, Li Cai, Xinlin Zhuang, Xuan Lin, Xin Lu, and Aimin Zhou. 2024. https://doi.org/10.18653/v1/2024.findings-acl.51 Are U a joke master? pun generation via multi-stage curriculum learning towards a humor LLM . In Findings of the Association for Computational Linguistics: ACL 2024, pages 878--890, Bangkok, Th...

    doi:10.18653/v1/2024.findings-acl.51 2024

  6. [6]

    Rose Laub Coser. 1959. Some social functions of laughter: A study of humor in a hospital setting. Human relations, 12(2):171--182

    work page 1959

  7. [7]

    Pawel Dybala, Michal Ptaszynski, Jacek Maciejewski, Mizuki Takahashi, Rafal Rzepka, and Kenji Araki. 2010. Multiagent system for joke generation: Humor and emotions combined in human-agent conversation. J. Ambient Intell. Smart Environ., 2(1):31–48

    work page 2010

  8. [8]

    Julen Etxaniz, Gorka Azkune, Aitor Soroa, Oier Lopez de Lacalle, and Mikel Artetxe. 2024. https://openreview.net/forum?id=QocjHRR31U Berta QA : How much do language models know about local culture? In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track

    work page 2024

Show all 40 references
  1. [9]

    Pawe Gryka, Kacper Grado \'n , Marek Koz owski, Mi osz Kuty a, and Artur Janicki. 2024. Detection of ai-generated emails-a case study. In Proceedings of the 19th International Conference on Availability, Reliability and Security, pages 1--8

    2024

  2. [10]

    William P Hampes. 1992. Relation between intimacy and humor. Psychological reports, 71(1):127--130

    1992

  3. [11]

    Bryan Anthony Hong and Ethel Ong. 2008. https://doi.org/10.1109/ISUC.2008.28 Generating punning riddles from examples . In 2008 Second International Symposium on Universal Communication, pages 347--352

    2008 doi

  4. [12]

    Ayesha Idrees, Saira Batool, and Rukhsana Kausar. 2020. Styles of humor and interpersonal relationships in university students. FWU Journal of Social Sciences, 14(4):57--67

    2020

  5. [13]

    Tonglin Jiang, Hao Li, and Yubo Hou. 2019. Cultural differences in humor perception, usage, and implications. Frontiers in psychology, 10:123

    2019

  6. [14]

    Marina Jovic and Salaheddine Mnasri. 2024. Evaluating ai-generated emails: A comparative efficiency analysis. World Journal of English Language, 14(2)

    2024

  7. [15]

    Zhuoran Lu, Sheshera Mysore, Tara Safavi, Jennifer Neville, Longqi Yang, and Mengting Wan. 2024. Corporate communication companion (ccc): An llm-empowered writing assistant for workplace social media. arXiv preprint arXiv:2405.04656

    2024 arXiv

  8. [16]

    Owen H Lynch. 2002. Humorous communication: Finding a place for humor in communication research. Communication theory, 12(4):423--445

    2002

  9. [17]

    Jessica Mesmer-Magnus, David J Glew, and Chockalingam Viswesvaran. 2012. A meta-analysis of positive humor in the workplace. Journal of Managerial Psychology, 27(2):155--190

    2012

  10. [18]

    AI Meta, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2

    2024 arXiv

  11. [19]

    John C Meyer. 2000. Humor as a double-edged sword: Four functions of humor in communication. Communication theory, 10(3):310--331

    2000

  12. [20]

    John C Meyer. 2015. Understanding humor through communication: Why be funny, anyway? Lexington Books

    2015

  13. [21]

    Rada Mihalcea and Carlo Strapparava. 2005. https://aclanthology.org/H05-1067 Making computers laugh: Investigations in automatic humor recognition . In Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing, pages...

    2005

  14. [22]

    John Morreall. 1983. Humor and emotion. American Philosophical Quarterly, 20(3):297--304

    1983

  15. [23]

    Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology, pages 1--22

    2023

  16. [24]

    Sahithya Ravi, Patrick Huber, Akshat Shrivastava, Aditya Sagar, Ahmed Aly, Vered Shwartz, and Arash Einolghozati. 2024. Small but funny: A feedback-driven approach to humor distillation. arXiv preprint arXiv:2402.18113

    2024 arXiv

  17. [25]

    Dmitry Scherbakov, Nina Hubig, Vinita Jansari, Alexander Bakumenko, and Leslie A Lenert. 2024. The emergence of large language models (llm) as a tool in literature reviews: an llm automated systematic review. arXiv preprint arXiv:2409.04600

    2024 arXiv

  18. [26]

    Leixian Shen, Haotian Li, Yun Wang, and Huamin Qu. 2024. From data to story: Towards automatic animated data video creation with llm-based multi-agent systems. arXiv preprint arXiv:2408.03876

    2024 arXiv

  19. [27]

    Oliviero Stock and Carlo Strapparava. 2005. https://doi.org/10.3115/1225753.1225782 HAHA cronym: A computational humor system . In Proceedings of the ACL Interactive Poster and Demonstration Sessions , pages 113--116, Ann Arbor, Michigan. Association for Computational Linguistics

    2005

  20. [29]

    Alexey Tikhonov and Pavel Shtykovskiy. 2024 b . Humor mechanics: Advancing humor generation with multistep reasoning. arXiv preprint arXiv:2405.07280

    2024 arXiv

  21. [30]

    Jana Vizmuller. 1980. Psychological reasons for using humor in a pedagogical setting. Canadian modern language review, 36(2):266--271

    1980

  22. [31]

    Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al. 2024. A survey on large language model based autonomous agents. Frontiers of Computer Science, 18(6):186345

    2024

  23. [32]

    Shan Wang. 2024. Investigating the potential of large language models for automated writing scoring. In 2024 5th International Conference on Education, Knowledge and Information Management (ICEKIM 2024), pages 1091--1098. Atlantis Press

    2024

  24. [33]

    Shih-Hung Wu, Yu-Feng Huang, and Tsz-Yeung Lau. 2024. Humour classification by fine-tuning llms: Cyut at clef 2024 joker lab subtask humour classification according to genre and technique. In Working Notes of the Conference and Labs of the Evaluation Forum (CLEF 2024). CEUR Wo...

    2024

  25. [34]

    Changrong Xiao, Wenxing Ma, Sean Xin Xu, Kunpeng Zhang, Yufang Wang, and Qi Fu. 2024. From automation to augmentation: Large language models elevating essay scoring landscape. arXiv preprint arXiv:2401.06431

    2024 arXiv

  26. [35]

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. 2024. Qwen2 technical report. arXiv preprint arXiv:2407.10671

    2024 arXiv

  27. [36]

    Diyi Yang, Alon Lavie, Chris Dyer, and Eduard Hovy. 2015. https://doi.org/10.18653/v1/D15-1284 Humor recognition and humor anchor extraction . In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2367--2376, Lisbon, Portugal. Associa...

    2015 doi

  28. [37]

    Renxian Zhang and Naishi Liu. 2014. https://doi.org/10.1145/2661829.2661997 Recognizing humor on twitter . In Proceedings of the 23rd ACM International Conference on Conference on Information and Knowledge Management, CIKM '14, page 889–898, New York, NY, USA. Association for ...

    2014

  29. [38]

    Shanshan Zhong, Zhongzhan Huang, Shanghua Gao, Wushao Wen, Liang Lin, Marinka Zitnik, and Pan Zhou. 2024. Let's think outside the box: Exploring leap-of-thought in large language models with creative humor generation. In Proceedings of the IEEE/CVF Conference on Computer Visio...

    2024

  30. [39]

    Hwang, Swabha Swayamdipta, and Maarten Sap

    Xuhui Zhou, Hao Zhu, Akhila Yerukola, Thomas Davidson, Jena D. Hwang, Swabha Swayamdipta, and Maarten Sap. 2023. https://doi.org/10.18653/v1/2023.findings-acl.392 COBRA frames: Contextual reasoning about effects and harms of offensive statements . In Findings of the Associatio...

    2023 doi

  31. [40]

    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...

  32. [41]

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