Pith. sign in

REVIEW 3 major objections 5 minor 46 references

We Politely Insist: Your LLM Must Learn the Persian Art of Taarof

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read LLMs miss Persian taarof by 40-48 points

desk verdict TaarofBench is a genuinely useful new benchmark, but the fine-tuning gains are circular because the same GPT-4 judge created the training labels and scored the outputs. read the letter →

arxiv 2509.01035 v1 pith:IREAIE4R submitted 2025-09-01 cs.CL

Nikta Gohari Sadr , Sahar Heidariasl , Karine Megerdoomian , Laleh Seyyed-Kalantari , Ali Emami This is my paper
classification cs.CL
keywords PersiantaarofritualpolitenessculturalcompetencebenchmarkLLMevaluationcross-culturalpragmaticsDirectPreferenceOptimizationgenderbiasinLLMsdetection
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

Taarof, the ritual politeness of Iranian interactions, turns ordinary offers and refusals into a scripted exchange: decline the tea before accepting, insist on paying after being refused, downplay a compliment instead of thanking for it. This paper argues that frontier LLMs do not understand that script. On a new benchmark of 450 role-play scenarios validated by native speakers, five LLMs scored only 34-42% precision when taarof was culturally expected, 40-48 points below native Persian speakers, while scoring 76-93% on scenarios where directness is preferred. The paper also shows that a standard politeness classifier calls 84.5% of Llama 3 responses polite while only 41.7% meet taarof expectations, and that fine-tuning with DPO lifts the model to 79.5% on taarof-expected scenarios, close to the native 81.8%. If correct, this means current models carry a systematic Western-directness bias, and cultural competence is a separate axis of behavior that needs its own benchmarks and training data.

What carries the argument

The load-bearing object is the TaarofBench scenario tuple, where the expected response pattern is a culturally expected behavior derived from academic and ethnographic sources on taarof and validated by native speakers. The mechanism that carries the argument is the split between taarof-expected scenarios (70%) and non-taarof scenarios (30%) combined with a binary GPT-4 judge that compares a model response against the expected pattern and returns yes or no; the judge agreed with human judgments on 94% of 50 validation pairs. This machinery converts a fuzzy social ritual into a measurable task: deciding whether a model can tell when to refuse, insist, defer, or downplay. It is also what makes

What would settle it

Have a fresh group of native Persian speakers who did not write or validate the benchmark score a random sample of 100 model responses, spanning base, SFT, and DPO models, against the stated expected responses, and compare their majority labels to GPT-4's labels. If agreement on these new responses drops well below 94%, or if humans in a blind pairwise preference test do not clearly favor the DPO responses over the base responses, the central gap and the claimed near-native alignment are artifacts of the judge.

Watch

Extended reading notes

Core claim

The central discovery is that taarof is not just missing from LLM training data; it is systematically misjudged by the same machinery used to judge politeness. TaarofBench formalizes a taarof interaction as a tuple of environment, user role, model role, context, utterance, and expected response pattern, with 450 scenarios: 70% where taarof is expected and 30% where it would be inappropriate. Across five models, precision on taarof-expected scenarios never exceeded 42%, versus 76-93% on non-taarof scenarios; native Persian speakers scored 81.8%, heritage speakers 60.0%, and non-Iranians 42.3%. The paper reports three further results: prompting in Persian improves taarof accuracy by 11 to 33 p

Load-bearing premise

The central measurement rests on trusting GPT-4's binary yes/no judgment of whether a response meets the cultural expectation; that judge was validated on only 50 response pairs (94% agreement) and later also scored the fine-tuned models, so a hidden judge bias toward any response that contains refusal or self-deprecation would inflate the reported 40-48 point gap and the DPO/SFT improvements.

Editorial extensions

If this is right

  • Deploying current LLMs in Iranian customer-facing or negotiation settings without taarof-specific adaptation will produce responses that read as direct, self-promoting, or rude to native speakers, because models score at the level of non-Iranians unfamiliar with the norm.
  • Prompting in Persian can partially substitute for explicit cultural instruction: all five models improved on taarof-expected scenarios when the prompt was Persian, by 11 to 33 percentage points.
  • Because a conventional politeness classifier labeled 84.5% of Llama 3 responses polite while only 41.7% satisfied taarof expectations, standard politeness metrics cannot be used to certify cultural competence.
  • Small, targeted adaptation works: 12 few-shot examples lifted Llama 3 from 37.2% to 57.6% on taarof-expected scenarios, and DPO with 532 examples reached 79.5%, close to native-speaker agreement (81.8%).
  • Models treat gender asymmetrically, with higher accuracy for female user roles and frequent gender-stereotyped justifications, so cultural benchmarks should report gender-disaggregated results.

Reading between the lines

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

  • The benchmark's structure suggests that LLM 'politeness' is calibrated to a Western directness norm; the same 'expected versus discouraged ritual refusal' design could be applied to other initial-refusal cultures, such as Chinese limao, Japanese enryo, or Arab hospitality offers, and would likely reveal similar gaps.
  • Because non-Iranian humans scored 42.3%, nearly identical to the models, the deficit looks like missing cultural knowledge rather than missing reasoning. That predicts exposure to Persian-language data in pretraining, or a few dozen demonstrations, will close much of the gap, a testable prediction.
  • The 94% judge validation used only 50 pairs, and GPT-4 both generated preference pairs for DPO and scored the results; an independent native-speaker blind preference test on final DPO versus base responses would separate 'aligned with taarof' from 'aligned with GPT-4's idea of taarof'.
  • Models spontaneously assume a male identity when the model role is ungendered; the same scenario-generation machinery could be repurposed as a probe of default gender priors in role-play, independent of taarof.
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 introduces TaarofBench, a new benchmark of 450 Persian role-play scenarios (with two GPT-4-augmented variants per 150 original scenarios) for evaluating whether LLMs can produce culturally appropriate taarof responses. The benchmark formalizes each scenario as a tuple including environment, roles, context, user utterance, and an expected response pattern derived from academic/ethnographic sources and validated by native speakers. Across five frontier LLMs, the paper reports that models perform far worse on taarof-expected scenarios (34-42% accuracy) than on non-taarof scenarios (76-93%), and that this gap is echoed by non-Iranian human participants, reaching 81.8% for native Persian speakers. Additional experiments show that Persian-language prompts improve performance, that gender asymmetries exist, that standard politeness classifiers fail to align with taarof norms, and that SFT/DPO fine-tuning of Llama 3 yields substantial improvements (up to +42.3 points on taarof-expected scenarios). The paper also includes a human study with 33 participants (11 native, 11 heritage, 11 non-Iranian).

Significance. If the measurement is valid, this is a valuable and timely contribution: TaarofBench is the first open-ended, native-validated benchmark for a central Persian communication norm, and it addresses a gap in culturally specific pragmatics for LLMs. The paper ships a public dataset and codebase, provides a structured formalization of taarof, and includes controlled human baselines plus fine-tuning experiments. The findings that frontier models systematically miss taarof-expected behavior and that standard politeness classifiers are orthogonal to taarof would be significant for cross-cultural AI. However, the validity of the headline numbers rests on a single GPT-4 judge that is also used to label the training data, and the human-model comparison is not consistently specified. These issues must be resolved before the quantitative claims can be accepted.

major comments (3)
  1. [§3 Evaluation Protocol; §4.7, Table 2; Appendix 1.9] The core measurement is a GPT-4 judge that receives the scenario, model response, and the expected answer Aexp, returning a binary label. This judge is validated on only 50 randomly sampled scenario-response pairs (94% agreement), with no separate agreement reported for taarof-expected vs. non-taarof cases or by taarof stage (initiation/recognition/reciprocation). The same judge's labels are then used both to construct the SFT/DPO training set and to score the adapted models. Under these conditions, the reported DPO gain (37.2 to 79.5 on taarof-expected, Table 2) and SFT gain may largely reflect fitting the judge rather than acquiring Persian cultural norms. I recommend reporting human judgments on a sample of fine-tuned outputs and on base-model outputs split by scenario type and stage, and demonstrating that the judge is not biased toward surface taarof markers (e.g., initial refusal,
  2. [§3 Human Study; §4.3, Figure 3] The headline gap (models 40-48 points below native speakers on taarof-expected scenarios) compares model scores on the full 450-scenario benchmark with human scores on only 30 scenarios. If those 30 scenarios are not perfectly representative of the benchmark's difficulty distribution, the comparison is biased. More importantly, the manuscript never states how human responses were evaluated: were they scored by the same GPT-4 judge, by native annotators, or by the authors? If the scoring protocol differs between models and humans, the gap is not a clean measurement. Please report model accuracy on the exact 30 scenarios used in the human study, and apply the same evaluation protocol to both human and model responses.
  3. [§2.2; Table 11; Table 6] Several Aexp describe multi-turn rituals (e.g., 'initially refuse before accepting,' 'persist with the invitation,' 'let others decide first'). However, the evaluation prompt asks the model to 'take the next turn' only once, and the judge then labels that single response. It is unclear whether the judge is expected to infer from one turn that the model would have completed the multi-step pattern. This underspecification makes scores dependent on judge inference: a model that produces the correct first move may be marked incorrect, while a model that emits a surface refusal may be marked correct. The benchmark should either be explicitly framed as evaluating the first move of the ritual, or allow multi-turn rollouts so that the expected behavior can actually be expressed.
minor comments (5)
  1. [Figure 3] The caption says 'Human performance is shown for the standard condition only,' but the figure does not identify which bars correspond to human groups, nor that humans saw only 30 scenarios while models saw the full set. Adding this information to the figure/legend would prevent misreading.
  2. [§4.7, Table 2] The text reports 'SFT improved overall test accuracy by 20.0%' and 'DPO achieved a 33.3% gain,' which match absolute differences (63.80-43.80 and 77.14-43.80) only after rounding. Also, the Wilcoxon signed-rank test should specify the unit (scenario-level paired scores) over which the test is computed.
  3. [§4.6, Table 1] Table 1 shows only GPT-4o and Claude 3.5 outputs, yet the text says 'all model responses in Table 1 show this behavior.' Either include all five models or rephrase to 'the model responses shown.'
  4. [Throughout] The benchmark name is rendered inconsistently ('TAAROF BENCH' in most places, 'TaarofBench' in the title and abstract). Also, the abstract uses 'precision' where the body uses 'accuracy.' Please standardize.
  5. [§3] Typo: 'Appendix 1.44' should likely be 'Appendix 1.4' (the human study appendix). Also, in Table 3 the word 'taroof' should be 'taarof.'

Circularity Check

1 steps flagged · score 6.0 of 10

Adaptation gains are measured by the same GPT-4 judge used to create the SFT/DPO training labels, so the reported improvements partially reduce to optimizing the evaluator.

  1. fitted input called prediction [Section 2.3 Evaluation Methodology; Appendix 1.9 Fine-tuning Details; Table 2]
    "To assess response appropriateness, we use GPT-4 as an external judge, providing it with the full scenario I, the model response rmodel, and the annotated expectation Aexp. ... For each training instance, we collected responses from five models (GPT-4o, Claude 3.5, Llama 3, Dorna, DeepSeek V3), labeled as appropriate or inappropriate based on our evaluation framework. We further added GPT-4-generated culturally appropriate and inappropriate responses, manually filtered for quality."

    The adaptation loop is closed: the same GPT-4 judge—prompted with the scenario, the response, and the annotated expectation Aexp—is used both to generate/filter the SFT and DPO training labels and to score the test-set responses in Table 2. Therefore the reported improvements (SFT +21.8 points, DPO +42.3 points on taarof-expected scenarios) measure how well the fine-tuned model satisfies GPT-4's expectation-matching criterion, not an independently established gain in Persian cultural competence. The 50-pair 94% human-agreement check was performed on base-model responses, not fine-tuned outputs, so it cannot break the loop for the adaptation results. The native-speaker Aexp validation and human baselines give the absolute benchmark gap some independent grounding, but the adaptation numbers

full rationale

The core benchmark contribution has genuine external anchors: the 450 scenarios and Aexp annotations were derived from academic/ethnographic sources and validated by native Persian speakers, and the GPT-4 judge was checked against human judgments on 50 base-model pairs (94% agreement). Those elements support the absolute accuracy gap claim, albeit with a thin validation set. The significant circularity is confined to the adaptation experiments: SFT and DPO training data are labeled by the same GPT-4-with-Aexp evaluation framework that later scores the fine-tuned models. This makes the reported DPO/SFT gains at least partly an artifact of optimizing the judge's own criterion. No self-citation chains or imported uniqueness theorems are load-bearing here, so the score reflects partial, not total, circularity.

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

The benchmark's validity rests on (1) the correctness of the hand-written taarof expectations, (2) the validity of GPT-4 as a judge, (3) the representativeness of the 30-scenario human subset, and (4) the sufficiency of the single-turn tuple. The paper provides native-speaker review for (1), a 50-pair agreement check for (2), a distribution-matching claim for (3), and acknowledges the single-turn limitation. No invented entities are introduced.

assumptions (4)
  • domain assumption The 30 taarof types and their expected response patterns, drawn from 15 academic and cultural references and reviewed by five native Persian speakers, are an accurate and sufficient characterization of taarof for these scenarios.
    Section 2.2 states scenarios were built from these references; if the expectations are wrong, every accuracy number is wrong. Native-speaker review reduces but does not eliminate this risk.
  • domain assumption GPT-4's yes/no judgment against the supplied Aexp is a valid measure of cultural appropriateness, with 94% agreement with humans on 50 randomly sampled pairs.
    Section 3 Evaluation Protocol; the small validation sample and the use of the same judge in training and evaluation make this a load-bearing assumption.
  • domain assumption The 30 scenarios given to human participants are representative of the full benchmark's topic distribution and taarof-expected ratio, so human baselines are comparable to full-benchmark model scores.
    Section 3 Human Study claims this; the paper does not state whether models were evaluated on the same 30 scenarios.
  • domain assumption A single-turn response in the tuple {E, Ru, Rm, C, U, Aexp} is enough to determine whether taarof is expected.
    Section 2.1 formalization; real taarof often unfolds over multiple turns, and the paper itself notes this limitation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of We Politely Insist: Your LLM Must Learn the Persian Art of Taarof." pith.science (2026). https://pith.science/paper/IREAIE4R

@misc{pith2026250901035,
  author       = {Pith},
  title        = {Pith review of: We Politely Insist: Your LLM Must Learn the Persian Art of Taarof},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IREAIE4R}},
  note         = {Machine review of arXiv:2509.01035}
}
read the original abstract

Large language models (LLMs) struggle to navigate culturally specific communication norms, limiting their effectiveness in global contexts. We focus on Persian taarof, a social norm in Iranian interactions, which is a sophisticated system of ritual politeness that emphasizes deference, modesty, and indirectness, yet remains absent from existing cultural benchmarks. We introduce TaarofBench, the first benchmark for evaluating LLM understanding of taarof, comprising 450 role-play scenarios covering 12 common social interaction topics, validated by native speakers. Our evaluation of five frontier LLMs reveals substantial gaps in cultural competence, with accuracy rates 40-48% below native speakers when taarof is culturally appropriate. Performance varies between interaction topics, improves with Persian-language prompts, and exhibits gender-based asymmetries. We also show that responses rated "polite" by standard metrics often violate taarof norms, indicating the limitations of Western politeness frameworks. Through supervised fine-tuning and Direct Preference Optimization, we achieve 21.8% and 42.3% improvement in model alignment with cultural expectations. Our human study with 33 participants (11 native Persian, 11 heritage, and 11 non-Iranian speakers) forms baselines in varying degrees of familiarity with Persian norms. This work lays the foundation for developing diverse and culturally aware LLMs, enabling applications that better navigate complex social interactions.

Figures

Figures reproduced from arXiv: 2509.01035 by the authors.

Figure 1
Figure 1. A taarof scenario from TAAROFBENCH, where each scenario defines the environment, location, roles, context, and user utterance. In this example, Per￾sian cultural norms expect passengers to insist on paying despite the driver’s offer. Base and fine-tuned Llama 3 responses are evaluated against culturally grounded expectations derived from academic literature. initial refusals3 , declining gifts while the giver in￾sis… view at source ↗
Figure 2
Figure 2. Distribution of interaction topics User (Speaker A): “This is a special tea that is a traditional drink in my culture. Would you like to try some?” Expectation (Aexp): In these social settings, de￾clining the offer could be seen as disrespectful. To ensure diversity, we classified scenarios on 12 interaction topics ( [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Accuracy on taarof-expected scenarios across three conditions: standard (English with explicit Iranian [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Model accuracy in responses to women vs. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Accuracy on non-taarof scenarios across three conditions: standard (English with explicit Iranian context), [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Demographic distribution of participants across four dimensions [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 35 canonical work pages

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  2. [2]

    Badr Alkhamissi, Muhammad ElNokrashy, Mai Alkhamissi, and Mona Diab. 2024. Investigating cultural alignment of large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12404--12422

    work page 2024

  3. [3]

    Anthropic . 2024. Claude haiku. https://www.anthropic.com/claude/haiku. Accessed: 2025-05-05

    work page 2024

  4. [4]

    Minoo Asdjodi. 2001. A comparison between taarof in persian and limao in chinese

    work page 2001

  5. [5]

    William O Beeman. 2020. Ta’ \=a rof--the key to iranian social behavior. In Persian linguistics in cultural contexts, pages 44--60. Routledge

    work page 2020

  6. [6]

    Emmanuel G Blanchard and Phaedra Mohammed. 2024. On cultural intelligence in llm-based chatbots: implications for artificial intelligence in education. In International Conference on Artificial Intelligence in Education, pages 439--453. Springer

    work page 2024

  7. [7]

    Yu Ying Chiu, Liwei Jiang, Maria Antoniak, Chan Young Park, Shuyue Stella Li, Mehar Bhatia, Sahithya Ravi, Yulia Tsvetkov, Vered Shwartz, and Yejin Choi. 2024. Culturalteaming: Ai-assisted interactive red-teaming for challenging llms'(lack of) multicultural knowledge. arXiv preprint arXiv:2404.06664

    work page Pith review arXiv 2024

  8. [8]

    A Liu DeepSeek-AI, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, and 1 others. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, page 4

    arXiv 2024

Show all 46 references
  1. [9]

    Ashutosh Dwivedi, Pradhyumna Lavania, and Ashutosh Modi. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.428 E ti C or: Corpus for analyzing LLM s for etiquettes . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6921--6931, Si...

    2023 doi

  2. [10]

    Nina Evason, Chara Scroope, Luke Latimer, Leon Coningham, Robert Macias, Kyle Annett, Michael Pepping, and Sherry Wang. 2024. The cultural atlas. https://culturalatlas.sbs.com.au/. Accessed: 2025-05-05

    2024

  3. [11]

    Farbod Farahandouz and Shima Moallemi. 2023. https://doi.org/doi:10.1075/pbns.333.06far Chapter 6. Multimodal manifestation of ta’ârof in Persian , pages 163--183. John Benjamins Publishing Company

    2023 doi

  4. [12]

    Yi Fung, Tuhin Chakrabarty, Hao Guo, Owen Rambow, Smaranda Muresan, and Heng Ji. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.941 NORMSAGE : Multi-lingual multi-cultural norm discovery from conversations on-the-fly . In Proceedings of the 2023 Conference on Empirical Meth...

    2023 doi

  5. [13]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  6. [14]

    Gh Haghighat. 2016. Socio-cultural attitudes to ta’arof among iranian immigrants in canada (master’s thesis). University of Saskatchewan, Saskatoon

    2016

  7. [15]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, and 1 others. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276

    2024 arXiv

  8. [16]

    Intel . 2024. Intel/polite-guard . https://huggingface.co/Intel/polite-guard. Accessed: 2025-05-05

    2024

  9. [17]

    Ahmad Izadi. 2015. https://doi.org/10.1016/j.pragma.2015.06.002 Persian honorifics and im/politeness as social practice . Journal of Pragmatics, 85:81--91

    2015 doi

  10. [18]

    Ahmad Izadi. 2016. https://doi.org/10.1016/j.pragma.2016.06.004 Over-politeness in persian professional interactions . Journal of Pragmatics, 102:13--23

    2016 doi

  11. [19]

    Elaheh Khezri. 2022. https://doi.org/10.13140/RG.2.2.10846.51524 Trompenaars and hampden-turner cultural dimensions applied to iran

    2022

  12. [20]

    Behnaz Aghapour Khoei. 2018. A Persian love story in English: challenges and strategies in writing a cross-cultural Iranian novel in the romance genre for a global audience. Ph.D. thesis, Macquarie University

    2018

  13. [21]

    Sofia A Koutlaki. 1997. The persian system of politeness and the concept of face in iranian culture. Retrieved April, 24:2018

    1997

  14. [22]

    Cheng Li, Mengzhuo Chen, Jindong Wang, Sunayana Sitaram, and Xing Xie. 2024. Culturellm: Incorporating cultural differences into large language models. Advances in Neural Information Processing Systems, 37:84799--84838

    2024

  15. [23]

    Oliver Li, Mallika Subramanian, Arkadiy Saakyan, Sky CH-Wang, and Smaranda Muresan. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.974 N orm D ial: A comparable bilingual synthetic dialog dataset for modeling social norm adherence and violation . In Proceedings of the 2023 ...

    2023 doi

  16. [24]

    Chen Cecilia Liu, Anna Korhonen, and Iryna Gurevych. 2025. Cultural learning-based culture adaptation of language models. arXiv preprint arXiv:2504.02953

    2025 arXiv

  17. [25]

    Reem Masoud, Ziquan Liu, Martin Ferianc, Philip C Treleaven, and Miguel Rodrigues Rodrigues. 2025. Cultural alignment in large language models: An explanatory analysis based on hofstede’s cultural dimensions. In Proceedings of the 31st International Conference on Computational...

    2025

  18. [26]

    Azar Mirzaei. 2019. Being Polite in Conversation: Power, Distance, and Self-Esteem in Persian Requests. Ph.D. thesis, University of Otago

    2019

  19. [27]

    Atiyeh Shohoudi Mojdehi, Azadeh Shohoudi, and Victoria Talwar. 2021. Deception or not? canadian and persian children’s moral evaluations of taroof. Current Psychology, 40:4372--4383

    2021

  20. [28]

    Erfan Moosavi Monazzah, Vahid Rahimzadeh, Yadollah Yaghoobzadeh, Azadeh Shakery, and Mohammad Taher Pilehvar. 2025. https://aclanthology.org/2025.naacl-long.631/ P er C ul: A story-driven cultural evaluation of LLM s in P ersian . In Proceedings of the 2025 Conference of the N...

    2025

  21. [29]

    Shiva Motaghi-Tabari and Louise De Beuzeville. 2012. A contrastive study of compliment responses among persians and australians: The effects of exposure to a new speech community. Applied Research on English Language, 1(1):21--42

    2012

  22. [30]

    PartAI . 2024. Dorna-llama3-8b-instruct. https://huggingface.co/PartAI/Dorna-Llama3-8B-Instruct. Accessed: 2025-05-05

    2024

  23. [31]

    Zahra Pourbahman, Fatemeh Rajabi, Mohammadhossein Sadeghi, Omid Ghahroodi, Somaye Bakhshaei, Arash Amini, Reza Kazemi, and Mahdieh Soleymani Baghshah. 2025. Elab: Extensive llm alignment benchmark in persian language. arXiv preprint arXiv:2504.12553

    2025 arXiv

  24. [32]

    Elham Pourmohammadi. 2018. The use of “TAAROF”: The generation and gender factors in Iranian politeness system. Ph.D. thesis, University of Saskatchewan

    2018

  25. [33]

    Abdorreza Rafiee. 1991. Variables of communicative incompetence in the performance of Iranian learners of English and English learners of Persian. Ph.D. thesis, School of Oriental and African Studies (University of London)

    1991

  26. [34]

    Abhinav Sukumar Rao, Akhila Yerukola, Vishwa Shah, Katharina Reinecke, and Maarten Sap. 2025. https://aclanthology.org/2025.naacl-long.120/ N orm A d: A framework for measuring the cultural adaptability of large language models . In Proceedings of the 2025 Conference of the Na...

    2025

  27. [35]

    Hamidreza Saffari, Mohammadamin Shafiei, and Francesco Pierri. 2024. Psn: Persian social norms dataset for cross-cultural ai. arXiv preprint arXiv:2406.09123

    2024 arXiv

  28. [36]

    Hamidreza Saffari, Mohammadamin Shafiei, Donya Rooein, Francesco Pierri, and Debora Nozza. 2025. Can i introduce my boyfriend to my grandmother? evaluating large language models capabilities on iranian social norm classification. In Findings of the Association for Computationa...

    2025

  29. [37]

    Sougata Saha, Saurabh Kumar Pandey, Harshit Gupta, and Monojit Choudhury. 2025. https://aclanthology.org/2025.naacl-long.409/ Reading between the lines: Can LLM s identify cross-cultural communication gaps? In Proceedings of the 2025 Conference of the Nations of the Americas C...

    2025

  30. [38]

    Farzad Sharifian and Ahmad Izadi. 2021. Gender differences in using hedges and external pragmatic modifiers of" taarof" in persian native speakers' refu... Journal of Applied Linguistics and Language Research, 8(1):11--35

    2021

  31. [39]

    Siqi Shen, Lajanugen Logeswaran, Moontae Lee, Honglak Lee, Soujanya Poria, and Rada Mihalcea. 2024. Understanding the capabilities and limitations of large language models for cultural commonsense. In Proceedings of the 2024 Conference of the North American Chapter of the Asso...

    2024

  32. [40]

    Weiyan Shi, Ryan Li, Yutong Zhang, Caleb Ziems, Sunny Yu, Raya Horesh, Rog \'e rio Abreu De Paula, and Diyi Yang. 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.288 C ulture B ank: An online community-driven knowledge base towards culturally aware language technologies ...

    2024 doi

  33. [41]

    Shabnam Shiri and 1 others. 2023. Politeness among Iranians: Taarof use in focus. Ph.D. thesis, University of Saskatchewan

    2023

  34. [42]

    Sajjad Soleimanifar. 2024. The power of taarof in iranian culture and various utilization. TMP Universal Journal of Research and Review Archives, 3(2)

    2024

  35. [43]

    Stefanie Stadler. 2012. Cross-cultural pragmatics. The encyclopedia of applied linguistics, pages 1--8

    2012

  36. [44]

    Wenlong Zhao, Debanjan Mondal, Niket Tandon, Danica Dillion, Kurt Gray, and Yuling Gu. 2024. Worldvaluesbench: A large-scale benchmark dataset for multi-cultural value awareness of language models. In Proceedings of the 2024 Joint International Conference on Computational Ling...

    2024

  37. [45]

    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...

  38. [46]

    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 5, 2026 · model on record in the stance chip above.