Exploring LLM-generated Culture-specific Affective Human-Robot Tactile Interaction

Qiaoqiao Ren; Tony Belpaeme

REVIEW 3 major objections 4 minor 1 cited by

Exploring LLM-generated Culture-specific Affective Human-Robot Tactile Interaction

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Large language models can generate text descriptions of affective touch that convey six of twelve emotions above chance under matched culture, and cultural mismatch worsens both decoding and appropriateness.

desk verdict A plausible but statistically underbuilt study of LLM-generated touch descriptions; the cross-cultural results are interesting, but pooled analysis and internal contradictions make the comparative claims shaky. read the letter →

arxiv 2507.22905 v1 pith:OQPS54YC submitted 2025-07-02 cs.HC

Qiaoqiao Ren , Tony Belpaeme This is my paper · ORCID

classification cs.HC

keywords affectivetouchhuman-robotinteractionlargelanguagemodelsemotiondecodingculturaldifferencestactileappropriatenessjudgment

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 asks whether large language models (GPT-3.5, GPT-4, GPT-4o) can generate text descriptions of affective touch that people from different cultures can read and decode as specific emotions, and whether those descriptions are judged socially appropriate for human-robot contact. Ninety participants (Chinese, Belgian, or culturally unspecified) imagined the described touches and made forced-choice emotion judgments plus appropriateness ratings, yielding 3,240 judgments. Under matched cultural conditions, six of twelve emotions—anger, fear, gratitude, love, surprise, and sympathy—were decoded significantly above chance, while self-focused emotions such as pride, embarrassment, and envy were not. Cultural mismatch lowered decoding accuracy and increased judgments of inappropriateness, and human-to-robot expressions were rated more appropriate than robot-to-human ones. The point of the work is to test whether LLMs could serve as a source of culturally adaptive affective touch guidance for robots.

What carries the argument

The carrying mechanism is a prompting-and-evaluation pipeline. A large language model is prompted as an encoder (human or robot) that must express one of twelve emotions solely through touch on the bare arm within a cultural context (Chinese, Belgian, or unspecified), producing ten distinct tactile behaviour descriptions that specify touch type, intensity, rhythm, and cultural rationale. Each description becomes a stimulus in a forced-choice emotion-decoding task with thirteen options, paired with an appropriateness judgment; the same descriptions are rated in matched and mismatched culture conditions and in both interaction directions. The load-bearing comparison is decoding accuracy against a 1/13 chance level, with the twelve emotion categories drawn from six basic emotions, three prosocial emotions, and three self-focused emotions.

What would settle it

Present the same 36 touch-behaviour descriptions as physical or video-rendered robot touches—rather than text alone—to matched and mismatched cultural groups and measure forced-choice decoding accuracy; if anger, fear, gratitude, love, surprise, and sympathy no longer exceed chance in the matched condition, the textual result does not generalise to embodied touch.

Watch

Extended reading notes

Core claim

The central discovery, on the paper's own terms, is that LLMs can generate culture-specific text descriptions of affective touch that carry decodable emotion signals comparable in breadth to human-to-human touch results: for matched-culture conditions, anger, fear, gratitude, love, surprise, and sympathy are recognised above chance, whereas self-focused emotions are not. The paper also finds that the direction of the interaction changes social judgment, with human-to-robot touch rated more appropriate than robot-to-human touch, and that mismatching the cultural context of the described touch to the decoder's culture reduces accuracy and raises inappropriateness ratings. These results are presented as evidence that LLMs can be used to generate culturally adaptive affective touch scripts for robots, while cautioning that aggressive, overly intimate, or ambiguous behaviours are exactly the ones judged inappropriate.

Load-bearing premise

All 3,240 judgments came from reading a text description and imagining the touch, not from feeling it, so the study assumes text-to-touch transference is strong enough that imagined touch predicts real interpersonal touch.

Editorial extensions

If this is right

Robots whose touch scripts are generated by LLMs can expect reliably decodable expressions of anger, fear, gratitude, love, surprise, and sympathy when the cultural context matches the user.
Self-focused states such as pride, embarrassment, and envy should not be entrusted to touch alone in LLM-generated interaction designs.
Deploying a touch script generated for one culture in another culture will measurably reduce emotion communication and increase perceived inappropriateness, so cultural context must be supplied at prompt time.
Robot-initiated affective touch will be judged more harshly than identical touch initiated by a human, pushing designers to set stricter appropriateness filters on robot-to-human expressions.
Behaviours that read as aggressive, overly intimate, or undecodable are the ones flagged inappropriate, giving a concrete safety criterion for filtering LLM output.

Reading between the lines

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

A testable extension the paper leaves implicit is whether the same decodability holds when the textual descriptions are executed by a physical robot arm; if it does, LLMs can replace hand-authored gesture libraries for affective touch.
The forced-choice design with a 'none of these' option may overstate confusion for ambiguous emotions, but it also makes the six above-chance results conservative relative to a two-alternative test.
Because all questionnaires were in English and participants imagined rather than felt the touch, the cultural-mismatch effect could be partly linguistic rather than tactile; translating the descriptions and adding haptic rendering would separate those factors.
The finding that self-focused emotions resist decoding may point to a general limit of the touch channel itself, since human-to-human studies show a similar pattern, rather than a limit specific to LLMs.

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.

Desk Editor's Note

read the letter

Quick take: this is a small empirical study with a real question—can LLMs generate culturally appropriate affective touch descriptions—and the matched-culture decoding results are plausible. The design (forced-choice decoding, matched/mismatched cultural groups, direction manipulation) is a reasonable extension of the authors' prior LLM-touch work. But the analysis has a load-bearing weakness: everything is pooled as if 3,240 responses were independent, when each participant contributed 36 responses and repeated measures are clustered. The binomial tests use 54 trials per emotion per group but only 18 participants, so marginal hits like surprise at 20.4% are fragile. The comparative claims—direction and cultural mismatch—are not actually tested with participant-level statistics; the mismatch effect is simply a count of separately significant tests (6 vs 1), which is not a direct comparison.

There are also internal contradictions the authors must fix. Section 3.1 says participants rated 54 stimuli, but the design says 36 and the total (3,240) matches 90×36. Section 3.2 says human-to-robot interactions were more associated with 'inappropriate' judgments, then says participants were more accepting of human-initiated touch, which is the opposite; Table 3 and the abstract support the latter. And the design says each of 18 series was rated by three participants, which only accounts for 54 participants, not 90. That last one might be a typo, but it matters for understanding the data structure.

On the positive side, the confusion matrices are genuinely informative, the finding that self-focused emotions (pride, embarrassment, envy) are hard to decode mirrors human-human touch literature, and the authors are honest about the text-to-touch proxy in the limitations. The paper would be much stronger with a mixed-effects model, a direct matched-vs-mismatched contrast, and a human-generated baseline or a softened claim.

Who is it for: HRI and affective computing researchers who care about culturally adaptive robot touch. It doesn't deserve a desk reject; the topic is timely and the core pattern is probably real. It needs a serious referee and a major revision before publication.

Referee Report

3 major / 4 minor

Summary. This manuscript reports an empirical study in which 90 participants (36 Chinese, 36 Belgian, 18 culturally unspecified) judged LLM-generated text descriptions of affective touch for 12 emotions, under three cultural prompt conditions (Chinese, Belgian, unspecified) and two interaction directions (robot-to-human and human-to-robot). The authors report that six emotions (anger, fear, gratitude, love, surprise, sympathy) are decoded above chance in matched cultural conditions, that self-focused emotions (embarrassment, pride, envy) are not, that human-to-robot touch is rated more appropriate than robot-to-human touch, and that cultural mismatch reduces decoding accuracy and increases the likelihood of inappropriate judgments.

Significance. If the findings survive a clustered re-analysis, this work makes a useful empirical contribution to human-robot interaction: it provides a transparent forced-choice protocol with a chance baseline for evaluating LLM-generated affective touch, compares three LLM families, and makes the stimulus-generation prompt publicly available on GitHub, all of which support reproducibility. The study also addresses a timely question about cultural adaptation in LLM-driven robots. The main weakness is that the current statistical tests do not account for the nested data structure (responses within participants and within stimulus series), so the headline quantitative claims are not yet statistically supported.

major comments (3)

[Section 3.1 and 3.2] The statistical analyses treat all 3,240 individual responses as independent observations. In the design, each participant rated 36 behaviours (3 per emotion) and each 36-behaviour series was rated by 3 participants (Section 2.3). Consequently, the 54 responses per emotion-group cell in Section 3.1 come from only 18 participants and 6 stimulus series, and the chi-square tests in Section 3.2 pool the same clustered data. This violates the independence assumption, and the reported p-values (e.g., surprise at 20.4%, padj<.05 in the None group; love at 22.2%, padj<.01 in Chi-Bel) are likely to change materially under a mixed-effects model with random intercepts for participants and series. Moreover, the matched-vs-mismatched comparison (6 of 12 vs 1 of 12 emotions decoded) is a comparison of counts of separately significant tests, not a statistical test of a difference. Please re-analyze using mixed-effects logistic regression or participant-level aggregation, and add direct group contrasts for decoding accuracy and appropriateness.
[Section 3.2, final paragraph] The reported direction of the receiver-identity effect contradicts Table 3 and the Conclusion. The text states that human-to-robot interactions were significantly more associated with inappropriate judgments, while robot-initiated expressions were more likely to be judged appropriate; however, Table 3 shows for every group that the robot-receiver row (human-to-robot) has a higher proportion of 'Appropriate' and a lower proportion of 'Inappropriate' than the human-receiver row (robot-to-human). The Conclusion correctly summarises the data as 'more acceptable when humans expressed emotion toward robots'. The Results text and the associated chi-square statistics need to be corrected to match the data.
[Section 3.1] There is a numerical inconsistency in the stimulus counts: Section 2.3 says each participant rated one series of 36 behaviours, but Section 3.1 states 'Each participant rated 54 stimuli' while reporting 3,240 total measurements. Since 90×36=3,240, the correct per-participant count is 36; the '54' likely refers to the number of responses per emotion per group (18 participants × 3 behaviours). Please correct this and verify all derived counts.

minor comments (4)

[Table 3] The column header 'Reciever' is misspelled; also clarify whether 'Receiver' refers to the target of the touch (human or robot), since this is central to interpreting the direction effect.
[Section 3.2] The chi-square statistics are reported with inconsistent p-value thresholds (e.g., χ²=106.0 with p<.001, χ²=82.4 with p<.01, χ²=124.0 with p<.01); given df=1, these values all correspond to p<.001. Please report exact p-values or use consistent thresholds.
[Section 3.2] Please state explicitly whether the 'Maybe' responses were excluded from the chi-square analyses; the text reports only 'appropriate' and 'inappropriate' counts.
[Section 3.4] There are several typos, including 'the participants’s attention' (should be 'participants' attention') and 'modelsociallyappropriatetouch' (missing spaces).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LLM output decoding is evaluated against independent human judgments; no fitted parameter, definitional equivalence, or load-bearing self-citation appears.

full rationale

The paper's central claims are empirical: LLM-generated text descriptions of touch were shown to participants, who attempted forced-choice decoding and judged appropriateness. The stimuli are generated from a fixed prompt (Section 2.2) and are not fitted to participant responses; the decoding rates (Section 3.1) and appropriateness distributions (Section 3.2) are measured, not derived from the prompt or from each other. No parameter is fitted and then renamed as a prediction, and no target quantity appears in the definition of the inputs. The cultural mismatch comparison is a between-subject empirical comparison (Chi-Chi vs. Chi-Bel, Bel-Bel vs. Bel-Chi), not a theorem or a self-citation chain. The two self-citations (refs. [9] and [13]) appear in the introduction as background motivation and are not used as evidence for the new results. The acknowledged limitations, such as text-only presentation and English questionnaires (Section 3.4), are external-validity concerns rather than definitional circularity. The pooled trial-level binomial and chi-square analyses may pose a statistical-inference risk, but that is a correctness concern, not a circularity concern.

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

The experimental claims rest on assumptions about what text-based prompts measure, about cultural labels, and about the forced-choice chance baseline. No fitted numerical parameters or new theoretical entities are introduced.

assumptions (4)

domain assumption Textual descriptions of touch are a valid proxy for actual tactile interaction.
The entire study asks participants to imagine the described interaction and judge emotion and appropriateness from text alone (Section 2.4). The authors acknowledge this as a limitation, and the results may not transfer to physical robot touch.
domain assumption Self-reported cultural background (Chinese, Belgian, unspecified) adequately indexes cultural touch norms.
Participants are grouped by cultural background with no validated measure of cultural orientation or touch norms, yet the matched and mismatched conditions rely on these labels (Section 2.3).
domain assumption The forced-choice menu with thirteen options makes 1/13 the correct chance baseline for decoding.
Decoding accuracy is tested against 1/13 chance with binomial tests (Section 3.1). This assumes participants choose uniformly at random when guessing, which is plausible but not verified.
domain assumption English-language questionnaires are adequate for Chinese and Belgian participants.
All materials were in English though English is not the native language of most participants; the authors list this as a possible limitation in Section 3.4.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Exploring LLM-generated Culture-specific Affective Human-Robot Tactile Interaction." pith.science (2026). https://pith.science/paper/OQPS54YC

@misc{pith2026250722905,
  author       = {Pith},
  title        = {Pith review of: Exploring LLM-generated Culture-specific Affective Human-Robot Tactile Interaction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OQPS54YC}},
  note         = {Machine review of arXiv:2507.22905}
}

read the original abstract

As large language models (LLMs) become increasingly integrated into robotic systems, their potential to generate socially and culturally appropriate affective touch remains largely unexplored. This study investigates whether LLMs-specifically GPT-3.5, GPT-4, and GPT-4o --can generate culturally adaptive tactile behaviours to convey emotions in human-robot interaction. We produced text based touch descriptions for 12 distinct emotions across three cultural contexts (Chinese, Belgian, and unspecified), and examined their interpretability in both robot-to-human and human-to-robot scenarios. A total of 90 participants (36 Chinese, 36 Belgian, and 18 culturally unspecified) evaluated these LLM-generated tactile behaviours for emotional decoding and perceived appropriateness. Results reveal that: (1) under matched cultural conditions, participants successfully decoded six out of twelve emotions-mainly socially oriented emotions such as love and Ekman emotions such as anger, however, self-focused emotions like pride and embarrassment were more difficult to interpret; (2) tactile behaviours were perceived as more appropriate when directed from human to robot than from robot to human, revealing an asymmetry in social expectations based on interaction roles; (3) behaviours interpreted as aggressive (e.g., anger), overly intimate (e.g., love), or emotionally ambiguous (i.e., not clearly decodable) were significantly more likely to be rated as inappropriate; and (4) cultural mismatches reduced decoding accuracy and increased the likelihood of behaviours being judged as inappropriate.

Figures

Figures reproduced from arXiv: 2507.22905 by the authors.

**Figure 2.** Emotions across cultural groups (top) and for misinterpreted or inappro [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

DRF: LLM-AGENT Dynamic Reputation Filtering Framework
cs.AI 2025-09 conditional novelty 4.0 of 10

DRF combines an LLM rating network, a reputation update rule, and a UCB-style selection strategy to filter low-quality LLM agents during multi-agent task execution, reporting improved pass@1 and lower simulated cost o...

Reference graph

Works this paper leans on

15 extracted references · 13 canonical work pages · cited by 1 Pith paper

[1]

Emotions: Investigating the vital role of tactile interaction

Xinyi Chen and Meng Ting Zhang. Emotions: Investigating the vital role of tactile interaction. In International Conference on Human-Computer Interaction, pages 326–344. Springer, 2024. 12 Q. Ren et al

work page 2024
[2]

Social robots on a global stage: establishing a role for culture during human–robot interaction.International Jour- nal of Social Robotics, 13(6):1307–1333, 2021

Velvetina Lim, Maki Rooksby, and Emily S Cross. Social robots on a global stage: establishing a role for culture during human–robot interaction.International Jour- nal of Social Robotics, 13(6):1307–1333, 2021

work page 2021
[3]

The science of interpersonal touch: an overview

Alberto Gallace and Charles Spence. The science of interpersonal touch: an overview. Neuroscience & Biobehavioral Reviews, 34(2):246–259, 2010

work page 2010
[4]

Affective interpersonal touch in close relationships: A cross-cultural perspective

Agnieszka Sorokowska, Supreet Saluja, Piotr Sorokowski, Tomasz Frąckowiak, Ma- ciej Karwowski, Toivo Aavik, Grace Akello, Charlotte Alm, Naumana Amjad, Afifa Anjum, et al. Affective interpersonal touch in close relationships: A cross-cultural perspective. Personality and Social Psychology Bulletin, 47(12):1705–1721, 2021

work page 2021
[5]

Generative artificial intelligence in robotic manipulation: A survey.arXiv preprint arXiv:2503.03464, 2025

Kun Zhang, Peng Yun, Jun Cen, Junhao Cai, Didi Zhu, Hangjie Yuan, Chao Zhao, Tao Feng, Michael Yu Wang, Qifeng Chen, et al. Generative artificial intelligence in robotic manipulation: A survey.arXiv preprint arXiv:2503.03464, 2025

arXiv 2025
[6]

Large language models for robotics: Op- portunities, challenges, and perspectives.Journal of Automation and Intelligence, 2024

Jiaqi Wang, Enze Shi, Huawen Hu, Chong Ma, Yiheng Liu, Xuhui Wang, Yincheng Yao, Xuan Liu, Bao Ge, and Shu Zhang. Large language models for robotics: Op- portunities, challenges, and perspectives.Journal of Automation and Intelligence, 2024

work page 2024
[7]

Survey of cultural awareness in language models: Text and beyond.arXiv preprint arXiv:2411.00860, 2024

Siddhesh Pawar, Junyeong Park, Jiho Jin, Arnav Arora, Junho Myung, Srishti Yadav, Faiz Ghifari Haznitrama, Inhwa Song, Alice Oh, and Isabelle Augenstein. Survey of cultural awareness in language models: Text and beyond.arXiv preprint arXiv:2411.00860, 2024

arXiv 2024
[8]

Generative expressive robot behaviors using large language models

Karthik Mahadevan, Jonathan Chien, Noah Brown, Zhuo Xu, Carolina Parada, Fei Xia, Andy Zeng, Leila Takayama, and Dorsa Sadigh. Generative expressive robot behaviors using large language models. In Proceedings of the 2024 ACM/IEEE International Conference on Human-Robot Interaction, pages 482–491, 2024

work page 2024

Show all 15 references

[9]

Touched by chatgpt: Using an llm to drive affective tactile interaction.arXiv preprint arXiv:2501.07224, 2025

Qiaoqiao Ren and Tony Belpaeme. Touched by chatgpt: Using an llm to drive affective tactile interaction.arXiv preprint arXiv:2501.07224, 2025

2025 arXiv
[10]

Towards an interactional approach to touch in social encounters

Asta Cekaite and Lorenza Mondada. Towards an interactional approach to touch in social encounters. InTouch in Social Interaction, pages 1–26. Routledge, 2020

2020
[11]

Social touch in human–robot in- teraction: Robot-initiated touches can induce positive responses without extensive prior bonding

Christian JAM Willemse and Jan BF Van Erp. Social touch in human–robot in- teraction: Robot-initiated touches can induce positive responses without extensive prior bonding. International journal of social robotics, 11(2):285–304, 2019

2019
[12]

The role of affective touch in human- robot interaction: Human intent and expectations in touching the haptic creature

Steve Yohanan and Karon E MacLean. The role of affective touch in human- robot interaction: Human intent and expectations in touching the haptic creature. International Journal of Social Robotics, 4:163–180, 2012

2012
[13]

Tactile interaction with social robots influences attitudes and behaviour

Qiaoqiao Ren and Tony Belpaeme. Tactile interaction with social robots influences attitudes and behaviour. International Journal of Social Robotics, 16(11):2297– 2317, 2024

2024
[14]

Touch communicates distinct emotions.Emotion, 6(3):528, 2006

Matthew J Hertenstein, Dacher Keltner, Betsy App, Brittany A Bulleit, and Ari- ane R Jaskolka. Touch communicates distinct emotions.Emotion, 6(3):528, 2006

2006
[15]

Integrat- ing visual context into language models for situated social conversation starters

Ruben Janssens, Pieter Wolfert, Thomas Demeester, and Tony Belpaeme. Integrat- ing visual context into language models for situated social conversation starters. IEEE Transactions on Affective Computing, 2024

2024

Pith tools

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

[1] [1]

Emotions: Investigating the vital role of tactile interaction

Xinyi Chen and Meng Ting Zhang. Emotions: Investigating the vital role of tactile interaction. In International Conference on Human-Computer Interaction, pages 326–344. Springer, 2024. 12 Q. Ren et al

work page 2024

[2] [2]

Social robots on a global stage: establishing a role for culture during human–robot interaction.International Jour- nal of Social Robotics, 13(6):1307–1333, 2021

Velvetina Lim, Maki Rooksby, and Emily S Cross. Social robots on a global stage: establishing a role for culture during human–robot interaction.International Jour- nal of Social Robotics, 13(6):1307–1333, 2021

work page 2021

[3] [3]

The science of interpersonal touch: an overview

Alberto Gallace and Charles Spence. The science of interpersonal touch: an overview. Neuroscience & Biobehavioral Reviews, 34(2):246–259, 2010

work page 2010

[4] [4]

Affective interpersonal touch in close relationships: A cross-cultural perspective

Agnieszka Sorokowska, Supreet Saluja, Piotr Sorokowski, Tomasz Frąckowiak, Ma- ciej Karwowski, Toivo Aavik, Grace Akello, Charlotte Alm, Naumana Amjad, Afifa Anjum, et al. Affective interpersonal touch in close relationships: A cross-cultural perspective. Personality and Social Psychology Bulletin, 47(12):1705–1721, 2021

work page 2021

[5] [5]

Generative artificial intelligence in robotic manipulation: A survey.arXiv preprint arXiv:2503.03464, 2025

Kun Zhang, Peng Yun, Jun Cen, Junhao Cai, Didi Zhu, Hangjie Yuan, Chao Zhao, Tao Feng, Michael Yu Wang, Qifeng Chen, et al. Generative artificial intelligence in robotic manipulation: A survey.arXiv preprint arXiv:2503.03464, 2025

arXiv 2025

[6] [6]

Large language models for robotics: Op- portunities, challenges, and perspectives.Journal of Automation and Intelligence, 2024

Jiaqi Wang, Enze Shi, Huawen Hu, Chong Ma, Yiheng Liu, Xuhui Wang, Yincheng Yao, Xuan Liu, Bao Ge, and Shu Zhang. Large language models for robotics: Op- portunities, challenges, and perspectives.Journal of Automation and Intelligence, 2024

work page 2024

[7] [7]

Survey of cultural awareness in language models: Text and beyond.arXiv preprint arXiv:2411.00860, 2024

Siddhesh Pawar, Junyeong Park, Jiho Jin, Arnav Arora, Junho Myung, Srishti Yadav, Faiz Ghifari Haznitrama, Inhwa Song, Alice Oh, and Isabelle Augenstein. Survey of cultural awareness in language models: Text and beyond.arXiv preprint arXiv:2411.00860, 2024

arXiv 2024

[8] [8]

Generative expressive robot behaviors using large language models

Karthik Mahadevan, Jonathan Chien, Noah Brown, Zhuo Xu, Carolina Parada, Fei Xia, Andy Zeng, Leila Takayama, and Dorsa Sadigh. Generative expressive robot behaviors using large language models. In Proceedings of the 2024 ACM/IEEE International Conference on Human-Robot Interaction, pages 482–491, 2024

work page 2024