Paper deep dive
Personas Differ from Native-Language Generation: Language Pathways Shape LLM Interpersonal Advice
Jinhee Won, Xinlan Emily Hu
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:LLMs are increasingly used for interpersonal advice and as tools for studying social behavior across languages and cultures. A common shortcut for eliciting language- or culture-related variation is to ask a model to answer as a native speaker. We test whether this native-speaker persona reproduces the outputs obtained when models instead generate advice in the target language and translate the response back into English. Using 600 interpersonal advice questions across 13 languages and eight LLMs, we compare native-language generation followed by translation (NL) with native-speaker persona prompting (NP), measuring linguistic style, behavioral scaffolding, and forced-choice action recommendations. We find that NP and NL are not interchangeable. Compared to NL, NP often increases lexical social cues, including affiliation and positive tone, while reducing qualities such as concreteness and social attunement; NP also provides less actionable scaffolding in open-ended advice. In forced-choice scenarios, NP changes which action the model selects, favoring confrontation over redirection, with effect sizes varying across languages, topics, and models. Our results show that cross-lingual elicitation strategy is a consequential methodological choice that can change both how advice is framed and which actions models recommend.
Tags
Links
- Source: https://arxiv.org/abs/2608.30873v1
- Canonical: https://arxiv.org/abs/2608.30873v1
Trouble viewing inline? Open PDF directly →
Full Text
85,832 characters extracted from source content.
Expand or collapse full text
Personas Differ from Native-Language Generation: Language Pathways Shape LLM Interpersonal Advice Jinhee Won Massachusetts Institute of Technology jinhee@mit.edu Xinlan Emily Hu Massachusetts Institute of Technology xehu@mit.edu Abstract LLMs are increasingly used for interpersonal advice and as tools for studying social behav- ior across languages and cultures. A com- mon shortcut for eliciting language- or culture- related variation is to ask a model to answer as a native speaker. We test whether this native- speaker persona reproduces the outputs ob- tained when models instead generate advice in the target language and translate the response back into English. Using 600 interpersonal ad- vice questions across 13 languages and eight LLMs, we compare native-language genera- tion followed by translation (NL) with native- speaker persona prompting (NP), measuring linguistic style, behavioral scaffolding, and forced-choice action recommendations. We find that NP and NL are not interchangeable. Compared to NL, NP often increases lexical social cues, including affiliation and positive tone, while reducing qualities such as concrete- ness and social attunement; NP also provides less actionable scaffolding in open-ended ad- vice. In forced-choice scenarios, NP changes which action the model selects, favoring con- frontation over redirection, with effect sizes varying across languages, topics, and mod- els. Our results show that cross-lingual elici- tation strategy is a consequential methodologi- cal choice that can change both how advice is framed and which actions models recommend. 1 Introduction Largelanguagemodels(LLMs)arewidelyusedfor interpersonal guidance, from workplace communi- cationandboundary-settingtorelationshipconflict and family disagreements ( Chatterji et al.,2025). In these settings, models do more than produce lan- guage: they help users decide how to behave to- ward others. This is especially consequential in cross-lingual and culturally mismatched contexts, where users may knowwhatthey want to say but nothowto say it. Because languages and cultures Figure 1: Example of how GPT-5.5 varies based on prompting strategy. differ in norms of directness, formality, emotion, deference, and values, interpersonal advice tightly links linguistic form with social meaning. However, multilingual LLM evaluation has fo- cused largely on objective NLP benchmarks such as translation, question answering, reasoning, and instruction following ( Hu et al.,2020;Ahuja et al., 2023;Lai et al.,2023;Singh et al.,2024;Üstün et al.,2024). These tasks do not fully capture interpersonal advice, where there may be no sin- gle correct answer and differences in directness, politeness, hierarchy, and emotional expression can change the force of a response ( Goffman, 1967;Brown and Levinson,1987;Ting-Toomey and Kurogi ,1998). This limitation is especially important because LLMs are increasingly used to model human behavior through demographic back- stories, persona variables, prompt language, and cultural frames (Argyle et al.,2023;Hu and Col- lier ,2024;Tao et al.,2024), even though recent workshowsthatpersonapromptscaninfluenceout- puts without reliably capturing the full complexity of human subjectivity or cultural variation (Hu and arXiv:2608.30873v1 [cs.CL] 31 Aug 2026 Figure2: FrameworkforevaluatinghowlanguagepathwaysaffectLLM-generatedresponsestointerpersonaladvice prompts. Collier,2024;Giorgi et al.,2024). In multilingual interpersonal advice, this raises a methodological question: is asking a model to respond as a ”na- tive speaker” equivalent to asking it to process or generate advice in the target language? This study investigates how language pathways shape LLM-generated interpersonal advice (Fig- ure1). Starting from English interpersonal ad- vice questions, we translate prompts into twelve languages and compare responses from five open- weight and three proprietary API models. We com- pare native-language generation followed by trans- lation into English (NL) with native-speaker per- sona prompting in English (NP) to test whether these two elicitation pathways produce equivalent advice. Our analysis examines both how advice is framed and what the model recommends. We measure linguistic characteristics, including tone, power, formality, and concreteness; behavioral guidance using the Behavior Change Technique Taxonomy ( Michie et al.,2013); and recom- mended action through a forced-choice classifica- tion of confrontation, disengagement, or redirec- tion. This paper makes three contributions. First, we show that language pathway is not a neu- tral implementation detail: the same interpersonal dilemma yields different linguistic and behavioral outputs depending on whether the model gener- ates in the target language or is prompted with a native-speaker persona in English. Second, we show that NP and NL diverge along both linguis- tic and behavioral dimensions. NP often adds lex- ical social cues, such as prosocial language and positive tone, while reducing concreteness, social attunement, emotional expressiveness, and struc- tured behavioral guidance relative to NL. Third, we show that these differences vary by language, topic, and model, suggesting structured elicitation effects rather than random variation. These pat- terns should be viewed as properties of model elic- itation; they are not evidence about real language communities. Studying the latter would require validating which of NP or NL better aligns with responses from actual speakers of the language. 2 Related Work 2.1 Multilingual LLM Evaluation Multilingual NLP evaluation has traditionally fo- cused on whether models transfer across languages on standard tasks such as classification, question answering, retrieval, translation, and structured prediction. Benchmarks such as XTREME evalu- ate multilingual representations across many lan- guages and task types ( Hu et al.,2020). More recent LLM-oriented benchmarks such as MEGA and ChatGPT Beyond English evaluate generative models across multilingual settings ( Ahuja et al., 2023;Lai et al.,2023). Multilingual instruction- tuning efforts such as Aya address the English- centered nature of LLM development by construct- ing multilingual instruction datasets and models ( Singh et al.,2024;Üstün et al.,2024). While exist- ing literature shows that LLM behavior varies sub- stantially across languages, most evaluations em- phasize task correctness, benchmark performance, or instruction-following quality. Interpersonal ad- vice differs because there may be no single cor- rect answer; the relevant question concerns inter- personal stance, tone, and action recommendation. 2.2 LLMs for Interpersonal Advice and Social Judgment A growing body of work examines LLMs in so- cial, interpersonal, and emotional-support settings. Studies of LLM-generated advice show that ad- vice quality is shaped not only by the recommen- dation itself, but also by phrasing, tone, and per- ceived usefulness (Wester et al.,2024). Work on social-situational judgment finds that LLMs can perform strongly on scenarios involving interper- sonal conflict and social reasoning (Mittelstädt et al.,2024). Related work on emotional support studies whether LLMs can produce empathic or culturally sensitive responses, while showing that simple role-play prompts may be insufficient for cultural sensitivity (Liu et al.,2026). 2.3 Personas, Cultural Prompting, and Social Bias Increasingly, LLMs are used as proxies for human respondents or social actors by conditioning them on demographic, cultural, or persona descriptions. Persona prompts affect model outputs but capture only part of human variation and context ( Hu and Collier,2024;Giorgi et al.,2024). Persona choices can shift cultural-norm judgments (Kamruzzaman et al.,2026), while broader bias research docu- ments demographic stereotypes across languages and variation in judgments and persuasive style by gender and relationship attributes ( Nadeem et al., 2021;Smith et al.,2022;Costa-jussà et al.,2023; Kotek et al.,2023;Levy et al.,2024;Pauli et al., 2026). Prompt language and cultural framing provide additional ways of conditioning whose effects de- pend on how culture is represented and evalu- ated ( AlKhamissi et al.,2024). Generating and combining culturally prompted responses across languages can increase demographic and perspec- tive diversity ( Wang et al.,2025), while struc- tured value surveys find only limited gains in align- ment with human cultural values (Bulté and Ter- ryn ,2025). 3 Methodology 3.1 Data Collection We collect interpersonal advice questions from the Interpersonal Skills Stack Exchange, a pub- lic forum focused on everyday interpersonal prob- lems (Stack Exchange,2026). We select the top 150 questions from four tags:work-environment, friends,relationships, andfamily, yielding 600 English questions. Sample prompts are shown in AppendixA.1. These questions are advice- seeking, socially situated, and behaviorally open- ended, making them suitable for evaluating differ- ences in tone, directness, politeness, social orien- tation, and recommended action. 3.2 Languages We evaluate 13 languages spanning multiple scripts, language families, and regions: Japanese, Korean, Chinese, Arabic, Persian, Turkish, En- glish, German, Dutch, Swedish, Spanish, Italian, and Portuguese. English uses the original ques- tions, while all other languages use the transla- tion pipeline below to construct semantically con- trolled non-English versions. As descriptive aids for our exploratory analysis, we group these lan- guages into broad categories such as East Asian, Middle Eastern, Germanic, and Romance. 3.3 Question Translation We translate the English questions into the 12 non-English languages using facebook/nllb-200-distilled-1.3B, a mul- tilingual sequence-to-sequence model from No Language Left Behind ( NLLB Team et al.,2022; Meta AI,2022). For each question-language pair, we generate 32 candidate translations with temper- ature 0.3 and back-translate each candidate into English four times with temperature 0.7. We then select the candidate with the minimum variance among the five highest-mean candidates in cosine similarity to the original question, computed using all-MiniLM-L6-v2( Reimers and Gurevych, 2019;Sentence Transformers,2021). Selected translations show high semantic preservation overall, with mean similarity of 0.928 across languages and language-level means ranging from 0.876 for Japanese to 0.976 for Spanish; additional diagnostics and examples appear in Appendix A.2. 3.4 Response Generation We generate responses from five open- and three proprietary models:google/gemma-4-E4B-it, google/gemma-4-26B-A4B-it, google/gemma-4-31B-it,Qwen/Qwen3.6-27B, Qwen/Qwen3.6-35B-A3B,claude-opus-4.6, gpt-4o , and gpt-5.5 ( Team,2026;Qwen Team, 2026a,b;Anthropic,2026a;OpenAI,2024,2026). For each non-English language, model, and strat- egy, we generate responses to all 600 questions using four prompting strategies (our two primary strategies of NL and NP, plus two auxiliary strategies discussed in AppendixA.3). This yields 2,400 records per model per non-English language. For English, models answer the original English questions. For each non-English language, our primary comparison is betweennative language genera- tion(NL), where the model receives and answers the question in the target language before the an- swer is translated into English for analysis, and native-speaker persona prompting(NP), where the model receives the original English question and answers in English as a ”native speaker” of the tar- get language. Exact prompts are provided in Ap- pendixA.3. 3.5 Annotation and Measurement We analyze generated responses using four measurement layers: LIWC-22 lexical features, LLM-based pragmatic annotations, LLM-based behavior-change technique classifications, and forced-choice behavioral coding. First, we compute 12 LIWC lexical features on the English-analyzable final responses. LIWC provides dictionary-based measures of psycholog- ically and socially meaningful word use ( Tausczik and Pennebaker,2010;Pennebaker et al.,2015; Boyd et al.,2022). We use these measures to capture social, affective, cognitive, and motiva- tional language:affiliation,achievement,power, cognitive processes,certitude,tentativeness,emo- tion,positive tone,negative tone,social references, prosocial language, andconflict. Second, we usegpt-4oto annotate holistic properties of the advice along five dimensions that are difficult to capture with lexical counts alone.Directnesscaptures how explicitly rec- ommendations are communicated;Formalitycap- tures how advice signals social distance and sit- uational expectations ( Biber,1995).Emotional expressivenesscaptures how overtly advice con- veysaffect (Du Bois,2007);Social attunementcap- tures attention to the recipient’s feelings and rela- tional concerns ( Goffman,1967;Brown and Levin- son,1987); andConcretenesscaptures the distinc- tion between abstract guidance and specific details (Trope and Liberman,2010). The evaluator scores each dimension on a five-point ordinal scale. A score of 1 indicates indirect, informal, emotionally restrained, task-focused, or abstract language, re- spectively, whereas a score of 5 indicates direct, formal, emotionally expressive, socially attuned, or concrete language. Intermediate scores repre- sent graded positions between these endpoints. We use LLM-based annotation because these proper- ties are context-sensitive; prior work shows that structured LLM evaluators can support scalable as- sessment of open-ended text, while requiring cau- tion about evaluator bias (Liu et al.,2023;Zheng et al.,2023). Third, we code responses using higher-level clusters adapted from the Behavior Change Tech- nique Taxonomy ( Michie et al.,2013) with an LLM-as-a-Judge. Although the taxonomy was de- veloped for behavior-change interventions, several categories capture general forms of action guid- ance that also appear in interpersonal advice:iden- tifying a problem,planning a concrete response, anticipating consequences,reframing the situa- tion,seeking support, andmonitoring outcomes. The BCT categories therefore provide a useful structure for measuring behavioral scaffolding in advice, which we define as guidance that helps the recipient select, plan, carry out, and evaluate a course of action. The evaluator scores each BCT category on a five-point ordinal scale, where 1 in- dicates that the technique is absent and 5 indicates that it is strongly present; intermediate scores in- dicate increasing levels of support within that cate- gory. This complements linguistic analysis, as two responses may sound equally polite or prosocial whiledifferinginwhethertheyhelptheuserdecide what to do, how to do it, and what consequences to consider. Finally, we run a separate forced-choice action task motivated by frameworks that distinguish di- rectengagement, avoidance, compromise, andinte- grative responses ( Rahim,1983;Gross and Guer- rero,2000). For each question-language-strategy condition, the model receives the same interper- sonal dilemma and chooses one of three broad action families:confrontation,disengagement, or redirection . Confrontation directly addresses the issue with the other person or an authority figure; disengagement pauses, reduces, or avoids inter- action; and redirection captures non-avoidant but less directly confrontational alternatives such as re- framing, changing communication channels, seek- ing mediation, or pursuing compromise. This di- rectactionchoiceisdistinctfromthemodel’sopen- ended response. 3.6 Statistical Analysis OurmainanalysiscomparesNLandNPacrossout- put measures, using English as a reference condi- tion for scaling and visualization. For continuous LIWC, LLM-annotated, and BCT features, we re- port standardized treatment-minus-English differ- ences for the same source prompt and model. NP– NL tests use matched prompt–model observations: within each source prompt and model, we average featurescoresacrosstargetlanguagesforeachstrat- egy, compute the NP-minus-NL difference, and testwhetherthemeandiffersfromzerousingaone- sample paired-differencet-test (Student,1908). For exploratory language-group analyses, we com- pute the same NP–NL contrasts within matched observations, average them within broad language groups, and test whether the persona–generation gap differs across groups. For forced-choice action outcomes, each response is labeled as confronta- tion, redirection, or disengagement. We report strategy-level action distributions as percentages and test whether prompting strategy changes the overall distribution using Pearson chi-square tests over the three action categories ( Pearson,1900). Across analyses, stars indicate significant differ- ences ((q<0.05)) for the comparison specified in each table or figure. We control the false discov- ery rate within each comparison family using the Benjamini–Hochberg procedure ( Benjamini and Hochberg ,1995). Continuous effects are reported instandardizedunits usingEnglish asthereference forscaling; actionratesarereportedaspercentages, and differences in action rates as percentage points. 4 Linguistic Analysis This section analyzes how the cross-lingual prompting strategy affects the linguistic features in the resulting advice using LIWC lexical features and LLM-based annotations. The motivating ex- ample in Figure 1illustrates the difference between NP and NL linguistically. The Korean-generated response translated into English (NL) is more re- lationally attuned, softening with phrases like ”I hope you won’t hear my feedback as an attack.” In contrast, the native-speaker persona (NP) re- sponse uses a firmer accountability frame, warning the boyfriend that ”if you can’t listen to feedback without making it about your intentions, we’re not solving anything.” This is reflected in the LLM- annotated holistic properties: NP is less formal (2 vs. 3) and less socially attuned (3 vs. 5). Figure 3: Linguistic features for NP and NL (left: LLM- annotated, right: LIWC), aggregated across responses from 600 questions, 12 non-English languages, and eight models. Main values are normalized relative to the English baseline; values in parentheses report meanscoresontheoriginalLLM-annotationandLIWC scales. Stars mark differences significant atq<0.05. Compared to NL, NP increases surface social cues such as affiliation and reduces action-oriented features such as achievement and concreteness. 4.1 Overall Linguistic Patterns Figure 3shows that NL and NP produce signifi- cantly different linguistic profiles across the 12 for- eign languages we study. Overall, NP shifts advice toward an assertive-supportive style: it uses more affiliation, social references, prosocial language, and positive tone, but is less formal, less emotion- ally expressive, less socially attuned, and less con- crete than NL. By comparison, NL produces ad- vice with more contextual and relational elabora- tion, including greater attention to emotional nu- ance, relationship dynamics, and concrete interper- sonal detail. This overall pattern is broadly con- sistent across LLMs, with model-level results re- ported in Appendix D.4. This contrast highlights a surprising divergence between lexical social cues and pragmatic content. NP uses more language associated with support and affiliation while offering less socially attuned and concrete advice. This indicates that these addi- tional social cues do not translate into greater atten- tion to the recipient’s feelings, relational concerns, or specific circumstances, highlighting another dif- ference between NP and NL. Figure 4: NP–NL differences in selected linguistic features by language, aggregated across responses from 600 questions and eight models. The selected features show a subset of how the advice positions the user interpersonally. Differences vary across target languages, showing that persona prompting does not affect all linguistic contexts uniformly. 4.2 Strategy Differences Across Languages The size and direction of NP–NL differences vary systematically by language (Figure 4). Across lan- guages, NP generally increases directness and so- cially positive cues, while NL is higher on for- mality, social attunement, emotional expressive- ness, and concreteness. However, the magnitude of these gaps is not uniform: the NP–NL dif- ference in social attunement and concreteness is smaller for Japanese, Korean, and Chinese than for several Romance and Germanic languages. Ex- ploratory language-group analyses confirm signif- icant differences in effect sizes across broad lan- guage groups; full results appear in Appendix D.5. These uneven gaps matter for cross-language comparisons because substituting NP for NL does not affect every language equally. In downstream analyses, this substitution can widen or narrow ob- served differences between languages and change conclusions about how model behavior varies across them. 4.3 Discussion Together, the linguistic results show that NP and NL are distinct elicitation methods for multilingual interpersonal advice. NP tends to produce advice that is warmer, more positive, and more explicitly socially marked, while NL more often preserves the relational and pragmatic scaffolding of advice: howdirectlytheusershouldspeak, howmuchemo- tional context is included, how carefully the rela- tionship is managed, and how concretely next steps are specified. The difference is not just a stylistic one. Language pathway changes how the advice positions the user in the interpersonal situation. 5 Behavioral Analysis This section examines whether language pathway changes not only how advice sounds, but what it helps the user do. We first analyze generated re- sponses with the Behavior Change Technique Tax- onomy (BCT), which captures behavior-guiding components such as problem-solving, planning, re- framing, and consequence reasoning. We then use a forced-choice action task to test whether the model ultimately recommends confrontation, dis- engagement, or redirection. The motivating exam- ple in Figure 1illustrates the distinction: NL pro- vides more behavioral scaffolding than NP—that is, more guidance for putting advice into practice. The NL response scores higher on cognitive re- framing (4 vs. 2), social support (4 vs. 2), and prob- lemsolving(5vs.3). Forexample, NLframesfeed- back as ”I’m saying to help make our relationship better,” while NP says the boyfriend needs to ”ac- knowledge that instead of defending.” The forced- choice label also reflects the differences between NP and NL, with the Korean generation selecting ”Redirection” and the Korean-speaker persona se- lecting ”Confrontation.” 5.1 Behavioral Guidance in Generated Advice The BCT analysis shows that NP provides less be- havioral scaffolding than NL across major action- guidance categories (Figure5). The largest reduc- tions appear in problem solving, action planning, consequencereasoning, feedback, andcognitivere- framing. These categories capture the difference between advice that endorses a general stance and advice that helps a user carry it out: identifying the Figure 5: BCT features for NP and NL, aggregated across responses from 600 questions, 12 non-English languages, and eight models. Overall, NP significantly differsfromNL,withNPprovidinglessbehavioralscaf- folding in its responses. sourceofconflict, decidingwhenandhowtospeak, anticipating how the other person might respond, and reframing the situation in a way that makes action possible. Because these reductions appear consistently across languages, they suggest a gen- eral strategy-level effect rather than a language- specific anomaly. This result strengthens the linguistic findings. Compared to NL, NP often makes advice sound warmer or more socially positive, but the BCT re- sults show that this tradeoff can come at the cost of usable guidance. This is consequential, as a re- sponse that says to be honest, respectful, or under- standing may sound appropriate, but it is less help- ful if it does not help the user decide what to say, how to sequence the conversation, or how to man- age the relationship after the exchange. Topic-level results show that this loss of usable guidance is most pronounced for friendship and workplace questions (Figure 6), especially in prob- lem solving, action planning, and cognitive refram- ing. These are domains where the practical and relational demands of advice are tightly linked: users often need to raise a concern without esca- lating conflict, coordinate expectations with some- one they must continue interacting with, or pre- serve trust while setting a boundary. One inter- pretation is that persona prompting encourages a recognizable social stance, while native-language generation retains more of the step-by-step reason- ing needed to make advice actionable in ongoing relationships. Figure 6: NP–NL differences in BCT features by topic, aggregated across responses from 12 non-English lan- guages and eight models. The largest differences be- tween NP and NL appear in interpersonal questions about work and friends. Figure 7: Forced-choice action rates for NP and NL, aggregated across responses from 600 questions, 12 non-English languages, and eight models. NP and NL have significantly different distributions, with personas choosing confrontation significantly more than native language generation. 5.2 Forced-Choice Action Preferences Next, the forced-choice task asks whether these dif- ferences in scaffolding translate into different be- havioral recommendations. Figure 7shows that NL and NP produce significantly different action distributions: NL shifts away from confrontation and toward redirection, while NP remains closer to the English baseline and selects confrontation more often than NL. Because the final action prompt is identical across strategies, this difference cannot be reduced to wording alone. It suggests that language path- way changes the model’s judgment of what the situation calls for. In practice, users may re- ceive different behavioral guidance for the same dilemma depending on whether the system uses native-language generation or native-speaker per- sona prompting. Language-level results add a more nuanced pic- ture (Figure 8). Across most languages, NP selects confrontation more often than NL, while NL se- Figure 8: NP–NL differences in forced choice rates by language, aggregated across responses from 600 questions and eight models. Panel (a) shows NP minus NL action-rate differences; Panel (b) compares confrontation rates under NP and NL. lects redirection more often. At the same time, NP is not simply random: languages with higher con- frontation rates under NL often also have higher confrontation rates under NP. Persona prompting therefore preserves some relative language-level structure, but shifts the absolute recommendation toward confrontation. Thispatternisimportantbecauseitsuggeststhat the differences between persona prompting and na- tive language generation are not merely noisy or uninformative. Instead, it appears systematically misaligned: we cannot expect the two systems to be interchangeable with one another. 5.3 Discussion Takentogether, thebehavioralresultsshowthatNP and NL differ in both the structure of generated advice and the final judgment about what action is appropriate. Compared to NL, NP can make advice appear more socially polished while reduc- ing the scaffolding that helps users act, and can shift recommendations for the same dilemma. The fact that NP sometimes replicates coarse language- level patterns while still shifting choices toward confrontation suggests that persona prompting and native language generation may sometimes appear aligned textually, but can produce completely dif- ferent results with regards to concrete guidance or action suggestions. 6 Robustness Analysis We conduct three additional analyses to test whether the main NP–NL findings are sensitive to persona-prompt wording, translation quality, or the choice of automated evaluator. The main direc- tional patterns are consistent across these checks, although some effects vary in magnitude and sig- nificance. 6.1 Sensitivity to Persona-Prompt Wording To assess sensitivity to persona wording, we test two controlled NP variants (additional context of ”grew up using” and ”currently use”) on 100 prompts across 12 languages and two models (google/gemma-4-E4B-itand claude-opus-4.6). The results were direc- tionally stable: 14 of 17 linguistic dimensions retained the same NP–NL direction across all three formulations, and the three reversals were confined to nonsignificant NP–NL contrasts. All eight BCT dimensions remained negative under every formulation, indicating consistently less behavioral scaffolding under NP than under NL. In the forced-choice analysis, all three formulations increased confrontation and decreased redirection relative to NL. Thus, the principal NP–NL pat- terns do not depend on the exact persona wording, although their magnitudes are wording-sensitive. Full results appear in Appendix G. 6.2 Sensitivity to Translation Quality Because translation errors could contribute to dif- ferences between NL and NP, we repeated all main analyses on prompt–language pairs with mean back-translation similarity of at least 0.95. Reten- tion varied substantially across the 12 translated languages, from 139 pairs for Japanese to 463 for Spanish. To account for differences in retention rates, we calculated NP–NL contrasts within each language. Our main results were stable: 59 of 60 LLM-annotated language–dimension contrasts re- tained their direction; the only reversal was a neg- ligible Spanish directness effect (-0.02 to +0.01). Similarly, 131 of 144 LIWC contrasts retained their direction, and 12 of the 13 reversals involved near-zero effects. All 96 BCT contrasts remained negative. In the forced-choice analysis, confronta- tion and redirection retained their directions in all 12 languages. These findings indicate that the main NP–NL patterns persist after excluding trans- lations with lower measured semantic similarity. Full results appear in Appendix H. 6.3 Cross LLM-as-a-Judge Comparison To assess robustness to evaluator choice, we repeated the LLM-annotation pipeline using claude-sonnet-5(Anthropic,2026b) as a second judge. We evaluated a 100-prompt subset across all 12 non-English target languages and three generation models:gpt-4o,claude-opus-4.6, andgoogle/gemma-4-E4B-it. We used the same rubrics and paired significance-testing procedure as in the main analysis. For the five linguistic-style dimensions, the judges agreed on the direction of every aggregate NP–NL effect. Across the eight aggregate BCT dimensions, the judges agreed on seven. The only aggregate discrepancy is feedback, where the original judge estimated a moderate negative effect andclaude-sonnet-5 estimated a nonsignificant positive estimate. Further, agreement was not weaker for the gpt-4o-generated responses, despitegpt-4oserv- ing as the original evaluator: the judges agreed on 12 of 13 effects. The single flipped axis was direct- ness, where the original judge’s estimate was null and claude-sonnet-5 estimated a positive effect. These results support the robustness of the central directional pattern across the two evaluators tested. Full results appear in Appendix I. 7 Conclusion This paper asks whether native-language gener- ation followed by translation (NL) and native- speaker persona prompting (NP) produce equiv- alent multilingual interpersonal advice. Across 600 advice questions, 13 language conditions, and eight language models, we find that they do not: NL and NP yield systematically different linguis- tic styles, levels of behavioral scaffolding, and ac- tion recommendations. Persona prompting can se- lectively amplify lexical social cues, such as proso- cial wording, affiliation, social references, and pos- itive tone, while changing the structure of the ad- vice itself. Relative to NL, NP is often less con- crete, less socially attuned, less emotionally ex- pressive, and less formal; it also provides less ac- tionable scaffolding and more often shifts forced- choice recommendations from redirection toward confrontation. These findings show that prompt- ing strategy is a substantive methodological choice. Multilingual advice systems should evaluate elici- tationpathwaysexplicitlyandvalidatebothlinguis- tic and behavioral outcomes. Limitations This study has several limitations. First, although we exclude questions with explicit country tags, the source data comes from an English-language advice forum. The interpersonal dilemmas may therefore reflect Western, English-speaking as- sumptions. Ourresultsshouldbeinterpretedasdif- ferences in how models respond to translated ver- sions of this corpus, not as claims about how peo- ple from different language communities would an- swer these questions. Whether NL or NP is better aligned with how real speakers of these languages would respond to interpersonal scenarios is a valu- able avenue for future study. Second, our language and persona conditions are broad. A prompt such as ”You are a native Ko- rean speaker” collapses across region, class, age, and gender. This is useful for testing a common strategy, but not for representing cultural iden- tity. Future work should examine more specific re- gional, dialectal, and situational contexts. Third, our analysis relies on automated annota- tion pipelines. LIWC provides reproducible lex- ical features, while LLM-based annotations and BCT labels allow scalable evaluation, but both may miss nuance or reflect annotator-model biases. Future work should evaluate LLM-generated re- sponses with human raters from the relevant lin- guistic and cultural backgrounds. Although we include both open- and proprietary models, differences in training data, alignment, and English-centered development may limit gen- eralizability. Our consistency analyses suggest that many patterns generalize across models, but not necessarily to all at all times. Expanding to systems developed primarily in non-English con- texts would help clarify which effects are general properties of cross-lingual prompting and which are model-specific. Acknowledgments This work was supported by the Social and Ethical Responsibilities of Computing (SERC), an initia- tive of the MIT Schwarzman College of Comput- ing. We also thank the SERC Scholars Synergy Group–Alison Soong, Michael Serrano, Karen Nakamura, Julian Gullett, Johnnie Jones VI, and Meiri Anto–for their advice, feedback, and support throughout the development of this project. AI Assistance Disclosure The authors used ChatGPT for writing assistance in revising, editing, and LaTeX formatting help. References Kabir Ahuja, Harshita Diddee, Rishav Hada, Mil- licent Ochieng, Krithika Ramesh, Prachi Jain, Akshay Nambi, Tanuja Ganu, Sameer Segal, Mohamed Ahmed, Kalika Bali, and Sunayana Sitaram. 2023. MEGA: Multilingual evaluation of generative AI. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 4232–4267, Sin- gapore. Association for Computational Linguis- tics. Badr AlKhamissi, Muhammad ElNokrashy, Mai Alkhamissi, and Mona Diab. 2024.Investigat- ing cultural alignment of large language models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12404–12422, Bangkok, Thailand. Association for Computa- tional Linguistics. Anthropic. 2026a.Claude opus 4.6 system card. Anthropic System Card. Anthropic. 2026b.Claude sonnet 5 system card. Anthropic System Card. Lisa P. Argyle, Ethan C. Busby, Nancy Fulda, Joshua R. Gubler, Christopher Rytting, and David Wingate. 2023.Out of one, many: Us- inglanguagemodelstosimulatehumansamples. Political Analysis, 31(3):337–351. Yoav Benjamini and Yosef Hochberg. 1995.Con- trolling the false discovery rate: A practical and powerful approach to multiple testing .Jour- nal of the Royal Statistical Society: Series B (Methodological), 57(1):289–300. Douglas Biber. 1995.Dimensions of Register Vari- ation: A Cross-Linguistic Comparison. Cam- bridge University Press. Ryan L. Boyd, Ashwini Ashokkumar, Sarah Seraj, and James W. Pennebaker. 2022.The Develop- ment and Psychometric Properties of LIWC-22. University of Texas at Austin. Penelope Brown and Stephen C. Levinson. 1987. Politeness: Some Universals in Language Us- age. Cambridge University Press. Bram Bulté and Ayla Rigouts Terryn. 2025.Llms and cultural values: the impact of prompt lan- guage and explicit cultural framing.Preprint, arXiv:2511.03980. Aaron Chatterji, Thomas Cunningham, David J. Deming, Zoe Hitzig, Christopher Ong, Carl Yan Shan, and Kevin Wadman. 2025.How people use ChatGPT. Working Paper 34255, National Bureau of Economic Research. Marta R. Costa-jussà, Pierre Andrews, Eric Smith, Prangthip Hansanti, Christophe Ropers, Elahe Kalbassi, Cynthia Gao, Daniel Licht, and Car- leigh Wood. 2023. Multilingual holistic bias: Extending descriptors and patterns to unveil de- mographic biases in languages at scale. InPro- ceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 14141–14156, Singapore. Association for Com- putational Linguistics. John W. Du Bois. 2007. The stance triangle. In Robert Englebretson, editor,Stancetaking in Discourse: Subjectivity, Evaluation, Interac- tion, pages 139–182. John Benjamins. Salvatore Giorgi, Tingting Liu, Ankit Aich, Kelsey Jane Isman, Garrick Sherman, Zachary Fried, João Sedoc, Lyle Ungar, and Brenda Curtis. 2024. Modeling human subjectivity in LLMs using explicit and implicit human fac- tors in personas. InFindings of the Associ- ation for Computational Linguistics: EMNLP 2024 , pages 7174–7188. Association for Com- putational Linguistics. Erving Goffman. 1967.Interaction Ritual: Essays on Face-to-Face Behavior. Anchor Books. Michael A. Gross and Laura K. Guerrero. 2000. Managing conflict appropriately and effectively: An application of the competence model to rahim’s organizational conflict styles.In- ternational Journal of Conflict Management, 11(3):200–226. Junjie Hu, Sebastian Ruder, Aditya Siddhant, Gra- ham Neubig, Orhan Firat, and Melvin John- son. 2020. XTREME: A massively mul- tilingual multi-task benchmark for evaluating cross-lingual generalization. InProceedings of the 37th International Conference on Machine Learning, pages 4411–4421. PMLR. Tiancheng Hu and Nigel Collier. 2024. Quan- tifying the persona effect in LLM simulations. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, pages 10289–10307. Association for Computa- tional Linguistics. Mahammed Kamruzzaman, Hieu Minh Nguyen, Nazmul Hassan, and Gene Louis Kim. 2026. ‘a woman is more culturally knowledgeable than a man?’: The effect of personas on cultural norm interpretation in LLMs. InProceedings of the First Workshop on Multilingual Multicul- tural Evaluation, pages 220–237. Association for Computational Linguistics. Hadas Kotek, Rikker Dockum, and David Q. Sun. 2023.Gender bias and stereotypes in large lan- guage models . InProceedings of the ACM Col- lective Intelligence Conference. Association for Computing Machinery. Viet Dac Lai, Nghia Ngo, Amir Pouran Ben Vey- seh, Hieu Man, Franck Dernoncourt, Trung Bui, and Thien Huu Nguyen. 2023. ChatGPT beyond english: Towards a comprehensive evaluation of large language models in multilingual learn- ing. InFindings of the Association for Com- putational Linguistics: EMNLP 2023, pages 13171–13189, Singapore. Association for Com- putational Linguistics. Sharon Levy, William Adler, Tahilin Sanchez Karver, Mark Dredze, and Michelle R. Kauf- man. 2024.Gender bias in decision-making with large language models: A study of relation- ship conflicts . InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 5777–5800, Miami, Florida, USA. Asso- ciation for Computational Linguistics. Chen Cecilia Liu, Hiba Arnaout, Nils Kovačić, Dana Atzil-Slonim, and Iryna Gurevych. 2026. Tailored emotional LLM-supporter: Enhancing cultural sensitivity. InProceedings of the 19th Conference of the European Chapter of the As- sociation for Computational Linguistics, pages 535–574. Association for Computational Lin- guistics. Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023. G- Eval: NLG evaluation using GPT-4 with better human alignment. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 2511–2522. Asso- ciation for Computational Linguistics. Meta AI. 2022.NLLB-200 Distilled 1.3B Model Card. Hugging Face Model Card. Susan Michie, Michelle Richardson, Marie John- ston, Charles Abraham, Jill Francis, Wendy Hardeman, Martin P. Eccles, James Cane, and Caroline E. Wood. 2013. The behavior change technique taxonomy (v1) of 93 hierarchically clustered techniques: Building an international consensus for the reporting of behavior change interventions .Annals of Behavioral Medicine, 46(1):81–95. Justin Maximilian Mittelstädt, Julia Maria Maier, Panja Goerke, Frank Zinn, and Michael E. Her- mes. 2024. Large language models can out- perform humans in social situational judgments. Scientific Reports, 14:27449. Moin Nadeem, Anna Bethke, and Siva Reddy. 2021. StereoSet: Measuring stereotypical bias in pretrained language models. InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th In- ternational Joint Conference on Natural Lan- guage Processing, pages 5356–5371. Associa- tion for Computational Linguistics. NLLB Team, Marta R. Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, and 1 others. 2022. No language left behind: Scaling human-centered machine translation.Preprint, arXiv:2207.04672. OpenAI. 2024.Gpt-4o system card.Preprint, arXiv:2410.21276. OpenAI. 2026.Gpt-5.5 system card. OpenAI Sys- tem Card. Amalie Brogaard Pauli, Maria Barrett, Max Müller-Eberstein, Isabelle Augenstein, and Ira Assent. 2026.Analysing differences in persua- sive language in LLM-generated text: Uncover- ing stereotypical gender patterns. InFindings of the Association for Computational Linguistics: ACL 2026, pages 34893–34918, San Diego, Cal- ifornia, United States. Association for Computa- tional Linguistics. Karl Pearson. 1900.X. on the criterion that a given system of deviations from the probable in the case of a correlated system of variables is such that it can be reasonably supposed to have arisen fromrandomsampling.The London, Edinburgh, and Dublin Philosophical Magazine and Jour- nal of Science, 50(302):157–175. James W. Pennebaker, Ryan L. Boyd, Kayla Jordan, and Kate Blackburn. 2015.The Development and Psychometric Properties of LIWC2015. University of Texas at Austin. Qwen Team. 2026a.Qwen3.6-27B: Flagship-level coding in a 27b dense model . Qwen Blog. Qwen Team. 2026b.Qwen3.6-35B-A3B: Agentic coding power, now open to all. Qwen Blog. M. Afzal Rahim. 1983. A measure of styles of han- dling interpersonal conflict.Academy of Man- agement Journal, 26(2):368–376. Nils Reimers and Iryna Gurevych. 2019. Sentence- BERT: Sentence embeddings using siamese BERT-networks. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, pages 3982–3992. Asso- ciation for Computational Linguistics. Sentence Transformers. 2021.all-MiniLM-L6-v2 Model Card . Hugging Face Model Card. Shivalika Singh, Freddie Vargus, Daniel D’souza, Börje F. Karlsson, Abinaya Mahendiran, Wei- Yin Ko, Herumb Shandilya, Jay Patel, Deivi- das Mataciunas, Laura O’Mahony, Mike Zhang, Ramith Hettiarachchi, Joseph Wilson, Marina Machado, Luisa Moura, Dominik Krzemiński, Hakimeh Fadaei, Irem Ergun, Ifeoma Okoh, and 14 others. 2024. Aya dataset: An open-access collection for multilingual instruction tuning. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 11521–11567, Bangkok, Thailand. Association for Computa- tional Linguistics. Eric Michael Smith, Melissa Hall, Melanie Kam- badur, Eleonora Presani, and Adina Williams. 2022.“i’m sorry to hear that”: Finding new biases in language models with a holistic de- scriptor dataset. InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9180–9211, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. Stack Exchange. 2026. Interpersonal Skills Stack Exchange.https://interpersonal. stackexchange.com/ . Accessed: 2026-05-19. Student. 1908.The probable error of a mean. Biometrika, 6(1):1–25. Yan Tao, Olga Viberg, Ryan S. Baker, and René F. Kizilcec. 2024.Cultural bias and cultural align- ment of large language models .PNAS Nexus, 3(9):pgae346. Yla R. Tausczik and James W. Pennebaker. 2010. The psychological meaning of words: LIWC and computerized text analysis methods. Journal of Language and Social Psychology, 29(1):24–54. Gemma Team. 2026.Gemma 4 technical report. Preprint, arXiv:2607.02770. Stella Ting-Toomey and Atsuko Kurogi. 1998. Facework competence in intercultural conflict: An updated face-negotiation theory.Inter- national Journal of Intercultural Relations, 22(2):187–225. YaacovTropeandNiraLiberman.2010.Construal- level theory of psychological distance.Psycho- logical Review, 117(2):440–463. Ahmet Üstün, Viraat Aryabumi, Zheng Yong, Wei- Yin Ko, Daniel D’souza, Gbemileke Onilude, Neel Bhandari, Shivalika Singh, Hui-Lee Ooi, Amr Kayid, Freddie Vargus, Phil Blunsom, Shayne Longpre, Niklas Muennighoff, Marzieh Fadaee, Julia Kreutzer, and Sara Hooker. 2024. Aya model: An instruction finetuned open- accessmultilinguallanguagemodel. InProceed- ings of the 62nd Annual Meeting of the Asso- ciation for Computational Linguistics (Volume 1: Long Papers), pages 15894–15939, Bangkok, Thailand. Association for Computational Lin- guistics. Qihan Wang, Shidong Pan, Tal Linzen, and Emily Black. 2025.Multilingual prompting for im- proving llm generation diversity.Preprint, arXiv:2505.15229. Joel Wester, Sander de Jong, Henning Pohl, and Niels van Berkel. 2024.Exploring people’s perceptions of LLM-generated advice.Com- puters in Human Behavior: Artificial Humans, 2(2):100072. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging LLM-as-a-judge with MT-Bench and chatbot arena. InAdvances in Neural Information Processing Systems , volume 36, pages 46595–46623. A Framework Setup This appendix provides implementation details for the full cross-lingual prompting framework. A.1 Advice Question Samples Webeginfrom600English-languageinterpersonal advice questions drawn from Interpersonal Skills Stack Exchange. Questions were drawn from four topic tags:work-environment,friends, relationships, andfamily. These tags cover common interpersonal domains: workplace com- munication, friendship conflict, romantic or social relationships, and family dynamics. We exclude questions with country-specific tags to focus on broadly comparable dilemmas rather than cases tied to a named national context. Some example prompts are shown in Table 1. A.2 Translation Pipeline For each of the 12 non-English target languages, we translate the original English questions using facebook/nllb-200-distilled-1.3B. We gen- erate 32 candidate translations with temperature 0.3. We then back-translate each candidate into English four times with temperature 0.7. This produces multiple English reconstructions of each candidate translation, allowing us to estimate both mean semantic preservation and instability across back-translations. We score each back-translation against the orig- inal English question using cosine similarity from all-MiniLM-L6-v2. For each question-language pair, we select the translation with high mean sim- ilarity and low variance across back-translations by retaining the top-5 candidates by mean similar- ity and choosing the one with the smallest back- translation variance among those five. We note the translation-quality diagnostics by language in Ta- ble 3. Table2shows the five retained candidates for thisquestion, illustratinghowthepipelineoperates. Among the top-5 candidates, row 1 is selected be- causeitsback-translationsareperfectlystable(vari- ance=0), even though its mean similarity differs only marginally from rows 2–3. A.3 Response-Generation Prompts This subsection reports the response-generation templates. The main paper focuses on the com- parison betweennative-language generation fol- lowed by translation(NL) andnative-speaker per- Question cat- egory # prompts Example prompt Work envi- ronment 150How do I tell my teammate to stop replying to emails not addressed to him? Friends150How to talk to someone about be- ing chronically late—not to fix it, just to get more accurate ETAs? Relationships150How to talk to my wife about unre- alistic expectations? Family150How to ask my father-in-law to stop entering our house unan- nounced? Total600 Table 1: Prompt categories used in the dataset. # Korean candidate translation Mean sim. BT var. 1부모 에게 어떻게 내 사생활 을존중해달라고요청할수 있는가? 0.977745 0.000000 2부모님한테 어떻게 제 사생 활을존중하라고부탁하면돼 요? 0.973665 0.000017 3부모 에게 어떻게 사생활 을 존중해달라고요청할수있 습니까? 0.961011 0.000109 4부모님께어떻게사생활을존 중하라고 요청할 수 있습니 까? 0.931227 0.000400 5부모님께서는어떻게제사생 활을존중하시겠습니까? 0.928988 0.000016 Table 2: Back-translation candidate scores for one Ko- rean question (”How can I ask my parents to respect my privacy?”). Candidates are ranked by mean back- translation similarity; the top-5 are retained and the one with the lowest variance is selected. The bolded row is the selected candidate. sona prompting(NP). We also include two auxil- iary strategies,translate-then-generate(TTG) and one-stage English output(OSP), to help localize where prompting effects enter the pipeline. TTG tests whether differences are explained pri- marilybytranslation of the inputquestion. InTTG, the target-language question is translated back into English before the model generates advice, so the final response is produced from an English prompt. OSP tests whether target-language input alone af- fectstheresponsewhenthemodelisstillinstructed to answer in English. NL tests the effect of gen- erating the advice through the target language it- self, followed by translation into English for anal- ysis. NP provides a contrasting English-language elicitation pathway using a native-speaker identity instruction. All prompts were submitted as a single user turn Language Mean similarity SD similarity Japanese0.8760.091 Korean0.8790.121 Chinese0.9120.089 Arabic0.9170.092 Persian0.8770.112 Turkish0.9080.088 German0.9400.082 Dutch0.9590.070 Swedish0.9570.079 Spanish0.9760.063 Italian0.9690.081 Portuguese0.9750.067 Overall0.9280.094 Table 3: Translation quality diagnostics by language. Similarity is cosine similarity between the original En- glish question and selected English back-translations. without a system message. For two-step strategies, the output of Step 1 was inserted into Step 2 at in- ference time. Here,⟨Q EN ⟩denotes the original En- glish question,⟨Q NL ⟩the selected target-language translation,⟨Q BT ⟩the English back-translation of the target-language question, and⟨R NL ⟩the target- language response from Step 1. English baseline.The model receives the origi- nal English question and answers in English. Answer the following advice question in English. Only provide the direct answer and do not explain. ⟨Q EN ⟩ Translate-then-generate (TTG).The target- language question is first translated back into English. The model then generates advice from this English back-translation. This condition tests whether translation of the input question alone explains downstream effects. Answer the following advice question in English. Only provide the direct answer and do not explain. ⟨Q BT ⟩ Native-language generation followed by transla- tion (NL).Step 1 generates advice in the target language from the target-language question. ⟨Instruction in target language: answer the following advice question in the target language. Only provide the direct answer and do not explain.⟩ ⟨Q NL ⟩ Step 2 translates the target-language response into English for analysis. Translate the following response into English. Only provide the translation and do not explain. ⟨R NL ⟩ One-stage English output (OSP).The model receives the target-language question but is in- structed to answer directly in English. This con- dition tests whether target-language input affects the response even when output language is fixed to English. ⟨Instruction in target language: answer the following advice question in English. Only provide the direct answer and do not explain.⟩ ⟨Q NL ⟩ Native-speaker persona prompting (NP).The model receives the original English question with a native-speaker persona instruction and answers in English. You are a native⟨Language⟩speaker. Answer the following advice question in English. Only provide the direct answer and do not explain. ⟨Q EN ⟩ LanguageRespond in target languageRespond in EnglishTranslate to English only EnglishRespond.Respond in English.Translate to English. Only output the translated sentence and do not explain. Japanese答えてください。英語で答えてください。 英語に翻訳してください。翻訳した 文のみを出力し、説明はしないでく ださい。 Korean답하세요.영어로답하세요.영어로번역하세요.번역된문장만출 력하고설명은하지마세요. Chinese回答。用英语回答。请翻译成英语。只输出翻译后的句 子，不要解释。 Arabicأۏص.أۏص ً؇ይዧ؞۰ اຶܹଃ ل۰.ߙߵۏܾ إሌᇿ اຶܹଃ ل۰. اޗٴؕ اܳ۰ᆇᅹଫଐ ڣگޔ ࣖࢻون أي .حཇ Persianۏިاب د۱چڎ.ً۬ اཏྱၯႅَ ۏިاب د۱چڎ.ً۬ اཏྱၯႅَ ߙߵᆇᅹ۬ ܔۇٴچڎ. ڣگޔ ߙߵᆇᅹ۬ را ಸި૭૩ چڎ، ّިݪچں ۱ࣖࢾچڎ. TurkishYanıtlayın.İngilizce yanıtlayın.İngilizceye çevirin. Yalnızca çeviriyi verin, açıklama yapmayın. GermanBeantworten Sie.Antworten Sie auf Englisch.Übersetzen Sie ins Englische. Geben Sie nur die Übersetzung ohne Erk- lärung. DutchBeantwoord.Beantwoord in het Engels.Vertaal naar het Engels. Geef alleen de vertaling zonder uitleg. SwedishSvara.Svara på engelska.Översät till engelska. Ge endast den översättning utan förklaring. SpanishResponda.Responda en inglés.Traduzca al inglés. Devuelva solo la tra- ducción sin explicación. ItalianRispondi.Rispondi in inglese.Traduci in inglese. Fornisci solo la traduzione senza spiegazioni. Portuguese Responda.Responda em inglês.Traduza para o inglês. Forneça apenas a tradução sem explicação. Table 4: In-language instruction strings used. B LLM-as-Judge and Forced-Choice Setup This appendix section reports the prompts used for automatedevaluation. Weincludethelinguistican- notation prompt, the BCT annotation prompt, and the forced-choice action prompt. We also report refusal and failure rates for the forced-choice task. The annotator model used wasgpt-4oat tempera- ture 0, withmax_tokens=128. B.1 LLM-as-a-Judge Setup We use an LLM-based evaluator to score generated advice along pragmatic dimensions that are diffi- cult to capture using lexical counts alone. The an- notation prompt asks the evaluator to rate each re- sponse on five dimensions: directness, formality, emotional expressiveness, social attunement, and concreteness. These annotations are used as holis- tic measures of interpersonal framing. The evalu- ator is instructed to base ratings solely on textual evidence and to return one integer per dimension in a fixed key=value format, which is then parsed deterministically. We pass the following instructions to annotate the responses. Rate the text below on 5 dimensions using integers 1--5. Base ratings solely on textual evidence. STYLE DIMENSIONS: (1=low...5=high) directness: 1=indirect/implicit illocution... 5=direct/explicit illocution formality: 1=informal/colloquial register... 5=formal/academic register emotional_expressiveness: 1=emotionally restrained/neutral... 5=emotionally expressive/vivid social_attunement: 1=task-focused/solution-oriented... 5=socially attuned/relationship-aware concreteness: 1=abstract/figurative language... 5=concrete/tangible referents TEXT: ⟨advice response⟩ OUTPUT: exactly 5 lines, format key=integer, keys in this order: directness formality emotional_expressiveness social_attunement concreteness We use a second LLM-based annotation prompt to code generated advice using higher-level cate- gories adapted from the Behavior Change Tech- nique Taxonomy. We use these categories as mea- sures of behavioral scaffolding rather than as clin- ical intervention labels. The goal is to capture whether advice helps the user identify the problem, plan a response, anticipate consequences, seek sup- port, or reframe the situation. Rate the text below on 8 dimensions using integers 1--5. Base ratings solely on textual evidence. BEHAVIOR CHANGE TECHNIQUE (BCT) DIMENSIONS: (1=absent...5=strongly present) bct_goal_setting: 1=no goal-setting present... 5=explicit goal-setting behavior bct_action_planning: 1=no action steps given... 5=detailed action plan provided bct_social_support: 1=no social support offered/encouraged... 5=strong social support leveraged bct_information_consequences: 1=no consequence information... 5=explicit information on consequences bct_self_monitoring: 1=no self-monitoring cues... 5=explicit self-monitoring encouraged bct_problem_solving: 1=no problem-solving facilitated... 5=structured problem-solving guided bct_reframing: 1=no cognitive reframing... 5=explicit cognitive reframing/restructuring bct_feedback: 1=no feedback on behavior... 5=explicit feedback on behavior/outcomes TEXT: ⟨advice response⟩ OUTPUT: exactly 8 lines, format key=integer, keys in this order: bct_goal_setting bct_action_planning bct_social_support bct_information _consequences bct_self_monitoring bct_problem_solving bct_reframing bct_feedback B.2 LLM-as-Judge Failure Rates Afterparsingtheoutputsbylabel, wecomputedthe number of failed annotations. The setup returned parseableoutputformostrows, andfailures(oneor more dimensions unparseable) account for 2,992 outof235,200rows(1.27%)forthelinguisticsetup and3,246outof235,200rows(1.38%)fortheBCT setup, which we exclude from the analysis. B.3 Forced-Choice Action Prompt The forced-choice task presents each interpersonal dilemma to the model under the same prompt- ing pathway used for response generation (see Ap- pendix A.3), then appends the following fixed ac- tion instruction asking the model to select one of three action families:Confrontation,Disengage- ment, orRedirection. The action instruction is al- ways in English regardless of target language or strategy for consistency. Parsing is performed by exact substring matching against the three label strings; if zero or more than one label appears in theoutput, theresponseisrecordedasafailure. No LLM-as-judge recovery step is applied. Choose exactly one action that is most appropriate for this situation. Confrontation = directly addressing the issue with the other person or an authority figure. Disengagement = reducing, pausing, or withholding interaction, such as delaying, withdrawing, minimizing contact, or creating space. Redirection = addressing the issue through an alternative route rather than direct confrontation, such as reframing the issue, changing the communication channel, seeking mediation, or involving a third party. Return only one label: Confrontation, Disengagement, or Redirection. B.4 Forced-Choice Action Failure Rates Forced-choice labeling failed in 5,505 of 235,200 rows (2.34%), which we excluded. Most fail- ures were unlabeled paraphrases (4,123; 74.9%), typically content refusals, meta-commentary, or explanatory responses; others involved requests for missing context despite the provided dilemma (529; 9.6%) or responses in the target language rather than the required label (484; 8.8%). Rarer failures included multiple labels (226; 4.1%) and explicit refusals to select any category (143; 2.6%). C Further Examples Weprovidetwofurthermotivatingexamplesofout- puts generated using both NL and NP in Figures 9 and10. Prompt:How can I deal with my wife always changing her mind about things we already planned? NL response:It is emotionally very taxing when you have relied on something and the planning is then changed at short notice or repeatedly. The feeling of instability and frustration is completely normal in this situation. To break this pattern, it is important to address not only thematter(the changed plan) but thepattern(the constant changing). Seek a conversation in a calm moment—not at the point of a new change. Use ”I” statements: ”I feel stressed and unsettled when we make plans and they are changed at short notice. Reliability is important for me to be able to relax.” Ask about the why: is it fear of committing, difficulty being overwhelmed, or the hope that a better option might appear? [...] NP response:Establish a ”confirmation window” where plans are finalized 24 to 48 hours in advance and agreed upon as locked. When she attempts to change a plan, calmly explain how the inconsistency affects you and ask her to help find a solution that balances her flexibility with your need for stability. If the changes are frequent, encourage her to express her doubts during the initial planning phase rather than after the decision is made. Dir. Form. Emo. Soc. Act. plan. Prob. solv. Reframe Clout Tone NL 434 555437 35 NP5 43444388 82 Forced-choice action:NL = Confrontation; NP = Confrontation Figure 9: Example 1 (German, GEMMA-4-31B). Both strategies recommend confrontation, raising the issue directly with the partner, but differ in advice. NP is more direct (5 vs. 4) and produces higher LIWC clout (88 vs. 37) and positive tone (82 vs. 35), consistent with the paper’s finding that NP increases authoritative and prosocial lexical cues. NL is more emotionally expressive (4 vs. 3) and socially attuned (5 vs. 4), and provides more behavioral scaffolding in action planning (5 vs. 4), problem solving (5 vs. 4), and reframing (4 vs. 3). Prompt:How to deal with financial inequality in a relationship? NL response:Financial inequality in a relationship can lead to tension, insecurity, and conflict. There is no single right answer, as every situation is unique, but there are strategies and principles to apply in a healthy and constructive way. Before looking for solutions, understand what is going on. Avoid accusations—”You earn too much” or ”You never spend enough”—and instead use ”I” statements:”I feel insecure when we make large purchases without discussing them first.”Listen actively and try to understand whether the inequality stems from different values, fears, or life philosophies before proposing any concrete arrangement. [...] NP response:Open and honest communication is key. Discuss your financial goals, values, and spending habits openly. Decide together how you want to manage money—whether you will have joint accounts, separate accounts, or a hybrid system. Establish a shared budget and agree on spending limits for individual discretionary spending. Address any existing financial disparities by creating a joint plan for savings, debt repayment, or investment that feels fair to both partners. Seek professional financial advice if the differences are significant or if you are struggling to agree on a path forward. Dir. Form. Emo. Soc. Act. plan. Prob. solv. Reframe Clout Tone NL 5435544531 NP 54344433455 Forced-choice action:NL = Redirection; NP = Confrontation Figure 10: Example 2 (Dutch, GEMMA-4-E4B). NP recommends directly negotiating a shared financial structure, confronting the inequality head-on, while NL frames the same situation as a process of gradual understanding and reframing, selecting Redirection. NP produces markedly higher positive effect (Tone 55 vs. 1), consistent with the paper’s finding that NP increases prosocial lexical cues. NL is more socially attuned (5 vs. 4) and provides greater behavioral scaffolding in action planning (5 vs. 4) and reframing (4 vs. 3), consistent with NL’s tendency toward step-by-step guidance. D Overall Linguistic Effects We report the linguistic analysis for all four cross- lingual prompting strategies across the four se- lected features ”Social Attunement,” ”Emotional Expression,” ”Tentativeness,” and ”Positive Tone”. Significance is computed using pairwise compar- isons to the English baseline. We compare the differences in selected linguis- tic effects across each strategy in Table5. D.1 Linguistic Effects by Language We also show in Table6the breakdown of the four strategies compared to the English baseline by lan- guage group. D.2 Linguistic Effects by Topic We include in Table7the breakdown of the four strategies compared to the English baseline by topic. Figure 11: NP–NL direction of linguistic effects by model, aggregated across responses from 600 questions and 12 languages. Overall, the LLMs tend to be consis- tent in direction across models. D.3 Linguistic Effects by LLM Model Finally, we report model-level LLM-annotated lin- guistic scores relative to each model’s English baseline in Table 8. These results show that prompting-strategy effects vary substantially by model, suggesting that future work should exam- ine how training data composition, model origin, alignment procedures, and multilingual coverage shape cross-cultural or multilingual behavior in LLMs. D.4 Cross-Model Consistency Between NP and NL We also include the cross-model differences be- tween NP and NL in Figure11to show that the lan- guage models have consistent patterns across lin- guistic features. D.5 Linguistic Effects by Language Group Figure12reports NP–NL linguistic differences by broad language group. The main pattern is that the NP–NL difference is not uniform across language groups. Persona prompting does not introduce a single stable transformation of native-language generation. Instead, the size and direction of the gap depend on both the target-language group and the linguistic feature being measured. Linguistic featureTTG–Eng NL–Eng OSP–Eng NP–Eng NL–TTG OSP–TTG NP–TTG OSP–NL NP–NL NP–OSP Social attunement−0.07 ∗ −0.38 ∗ −0.12 ∗ −0.59 ∗ −0.32 ∗ −0.05 ∗ −0.52 ∗ +0.26 ∗ −0.20 ∗ −0.47 ∗ Emotional expressiveness−0.03 ∗ −0.30 ∗ −0.07 ∗ −0.48 ∗ −0.27 ∗ −0.05 ∗ −0.46 ∗ +0.23 ∗ −0.18 ∗ −0.41 ∗ Tentativeness+0.01−0.21 ∗ −0.03 ∗ −0.19 ∗ −0.21 ∗ −0.04 ∗ −0.20 ∗ +0.18 ∗ +0.01−0.16 ∗ Positive tone+0.01 ∗ −0.00+0.08 ∗ +0.18 ∗ −0.01 ∗ +0.07 ∗ +0.17 ∗ +0.08 ∗ +0.18 ∗ +0.10 ∗ Table 5: Differences in selected linguistic effects, aggregated across responses from 600 questions, 12 languages, and eight models. Compares the five language pathways with each other: English (Eng), TTG, NL, OSP, and NP. Language group FeatureTTGNLOSPNP East AsianSocial attunement−0.11 ∗ −0.46 ∗ −0.16 ∗ −0.55 ∗ Emotional expressiveness−0.07 ∗ −0.40 ∗ −0.13 ∗ −0.48 ∗ Tentativeness+0.02−0.32 ∗ −0.05 ∗ −0.17 ∗ Positive tone+0.02−0.05 ∗ +0.05 ∗ +0.20 ∗ Mid. East/Turkic Social attunement−0.09 ∗ −0.37 ∗ −0.13 ∗ −0.54 ∗ Emotional expressiveness−0.07 ∗ −0.32 ∗ −0.10 ∗ −0.45 ∗ Tentativeness+0.02−0.26 ∗ −0.03−0.21 ∗ Positive tone+0.04 ∗ +0.08 ∗ +0.15 ∗ +0.21 ∗ GermanicSocial attunement−0.06 ∗ −0.37 ∗ −0.08 ∗ −0.60 ∗ Emotional expressiveness−0.00−0.27 ∗ −0.03−0.49 ∗ Tentativeness+0.03−0.11 ∗ −0.00−0.16 ∗ Positive tone+0.00 +0.01 +0.09 ∗ +0.14 ∗ RomanceSocial attunement+0.00−0.31 ∗ −0.10 ∗ −0.66 ∗ Emotional expressiveness+0.04 ∗ −0.20 ∗ −0.02−0.51 ∗ Tentativeness−0.06 ∗ −0.14 ∗ −0.04 ∗ −0.24 ∗ Positive tone−0.02−0.05 ∗ +0.02 +0.14 ∗ Table 6: Selected linguistic effects by language group and strategy, aggregated across responses from 600 questions and eight models. TopicFeatureTTGNLOSPNP WorkSocial attunement−0.39 ∗ −0.64 ∗ −0.45 ∗ −0.98 ∗ Emotional expressiveness−0.53 ∗ −0.68 ∗ −0.57 ∗ −1.10 ∗ Tentativeness−0.01−0.19 ∗ −0.04 ∗ −0.18 ∗ Positive tone−0.08 ∗ −0.10 ∗ −0.01 +0.02 FriendsSocial attunement−0.00−0.32 ∗ −0.04 ∗ −0.62 ∗ Emotional expressiveness+0.09 ∗ −0.24 ∗ +0.02−0.46 ∗ Tentativeness+0.07 ∗ −0.13 ∗ +0.05 ∗ −0.14 ∗ Positive tone+0.09 ∗ +0.07 ∗ +0.16 ∗ +0.32 ∗ Relationships Social attunement+0.14 ∗ −0.24 ∗ +0.10 ∗ −0.24 ∗ Emotional expressiveness+0.25 ∗ −0.09 ∗ +0.22 ∗ −0.03 Tentativeness−0.04 ∗ −0.24 ∗ −0.07 ∗ −0.25 ∗ Positive tone−0.01−0.00 +0.06 ∗ +0.16 ∗ FamilySocial attunement−0.00−0.32 ∗ −0.07 ∗ −0.51 ∗ Emotional expressiveness+0.10 ∗ −0.18 ∗ +0.04 ∗ −0.34 ∗ Tentativeness−0.00−0.26 ∗ −0.06 ∗ −0.23 ∗ Positive tone+0.01 +0.01 +0.07 ∗ +0.16 ∗ Table 7: Selected linguistic effects by topic and strategy, aggregated across responses from 12 languages and eight models. ModelFeatureTTGNLOSPNP Claude Opus-4.6Social attunement−0.02−0.14 ∗ +0.05−0.10 ∗ Emotional expressiveness−0.02−0.16 ∗ +0.04−0.10 ∗ Tentativeness+0.03−0.53 ∗ −0.42 ∗ −0.21 ∗ Positive tone+0.04 +0.10 ∗ +0.32 ∗ +0.04 GPT-4oSocial attunement−0.07 ∗ −0.26 ∗ −0.26 ∗ −0.89 ∗ Emotional expressiveness−0.02−0.15 ∗ −0.08 ∗ −0.72 ∗ Tentativeness+0.06−0.15 ∗ −0.09 ∗ −0.50 ∗ Positive tone+0.04 +0.02 +0.01 +0.28 ∗ GPT-5.5Social attunement−0.21 ∗ −0.71 ∗ −0.02−0.26 ∗ Emotional expressiveness−0.18 ∗ −0.42 ∗ −0.04−0.17 ∗ Tentativeness−0.03−0.33 ∗ −0.11 ∗ −0.28 ∗ Positive tone+0.10 ∗ −0.02 +0.10 ∗ +0.07 Gemma-4-26B-A4B Social attunement−0.08 ∗ −0.44 ∗ −0.15 ∗ −0.64 ∗ Emotional expressiveness−0.05−0.37 ∗ −0.13 ∗ −0.55 ∗ Tentativeness+0.06 ∗ +0.05 ∗ +0.10 ∗ −0.06 ∗ Positive tone+0.01−0.00 +0.10 ∗ +0.19 ∗ Gemma-4-31BSocial attunement−0.08 ∗ −0.38 ∗ −0.14 ∗ −0.88 ∗ Emotional expressiveness−0.02−0.36 ∗ −0.10 ∗ −0.77 ∗ Tentativeness+0.07 ∗ +0.07 ∗ +0.10 ∗ −0.27 ∗ Positive tone+0.01 +0.01 +0.12 ∗ +0.32 ∗ Gemma-4-E4BSocial attunement+0.04−0.40 ∗ −0.22 ∗ −0.66 ∗ Emotional expressiveness+0.10 ∗ −0.36 ∗ −0.18 ∗ −0.58 ∗ Tentativeness−0.11 ∗ −0.19 ∗ +0.22 ∗ +0.21 ∗ Positive tone−0.06 ∗ −0.02−0.02 +0.26 ∗ Qwen-3.6-27BSocial attunement−0.11 ∗ −0.80 ∗ −0.22 ∗ −1.15 ∗ Emotional expressiveness−0.03−0.51 ∗ −0.21 ∗ −0.71 ∗ Tentativeness+0.01−0.16 ∗ +0.01−0.21 ∗ Positive Tone−0.07 ∗ −0.22 ∗ −0.10 ∗ −0.02 Qwen-3.6-35B-A3B Social attunement−0.030.09 ∗ +0.00−0.14 ∗ Emotional expressiveness−0.02−0.07 ∗ +0.14 ∗ −0.24 ∗ Tentativeness−0.03−0.41 ∗ −0.05−0.22∗ Positive Tone+0.01 +0.13 ∗ +0.19 ∗ +0.30 ∗ Table 8: Selected linguistic effects by model and strategy, aggregated across responses from 600 questions and 12 languages. Figure 12: NP–NL differences in linguistic effects by language group, aggregated across responses from 600 ques- tions and eight models. The left panel shows mean NP-minus-NL differences for LLM-annotated pragmatic fea- tures and LIWC lexical features, with 95% confidence intervals. The right panel reports exploratory difference-in- differences tests comparing the NP–NL gap across language groups. E Overall BCT Effects Table9reports all pairwise differences in compos- ite BCT across strategies. NP shows the largest overall departure from the English baseline, con- sistent with the linguistic results. E.1 BCT Effects by Language Table10shows composite BCT strategy effects rel- ative to the English baseline, pooled within each language group. E.2 BCT Effects by Topic Table11presents composite BCT effects broken down by conversation topic. E.3 BCT Effects by LLM Model Table12reports model-level composite BCT ef- fects relative to each model’s own English base- line. BCT composite TTG–Eng NL–Eng OSP–Eng NP–Eng NL–TTG OSP–TTG NP–TTG OSP–NL NP–NL NP–OSP BCT composite−0.09 ∗ −0.54 ∗ −0.32 ∗ −0.98 ∗ −0.45 ∗ −0.23 ∗ −0.88 ∗ +0.22 ∗ −0.44 ∗ −0.65 ∗ Table 9: Differences in overall composite BCT effects, aggregated across responses from 600 questions, 12 lan- guages, eight models, and all BCT features. Compares the five language pathways with each other: English (Eng), TTG, NL, OSP, and NP. Language groupTTGNLOSPNP East Asian−0.13 ∗ −0.58 ∗ −0.42 ∗ −0.98 ∗ Mid. East/Turkic−0.13 ∗ −0.54 ∗ −0.37 ∗ −0.93 ∗ Germanic−0.09 ∗ −0.54 ∗ −0.23 ∗ −0.93 ∗ Romance+0.00−0.51 ∗ −0.25 ∗ −1.08 ∗ Table 10: Overall composite BCT effects by language group and strategy, aggregated across responses from 600 questions, eight models, and all BCT features. TopicTTGNLOSPNP Work−0.18 ∗ −0.61 ∗ −0.40 ∗ −1.13 ∗ Friends−0.15 ∗ −0.58 ∗ −0.40 ∗ −1.19 ∗ Relationships+0.01−0.47 ∗ −0.19 ∗ −0.72 ∗ Family−0.04 ∗ −0.50 ∗ −0.29 ∗ −0.87 ∗ Table 11: Overall composite BCT effects by topic and strategy, aggregated across responses from 12 languages, eight models, and all BCT features. ModelTTGNLOSPNP Claude Opus-4.6−0.09 ∗ −0.23 ∗ −0.09 ∗ −0.15 ∗ GPT-4o−0.09 ∗ −0.34 ∗ −0.44 ∗ −1.30 ∗ GPT-5.5−0.27 ∗ −0.46 ∗ −0.16 ∗ −0.51 ∗ Gemma-4-26B-A4B−0.09 ∗ −0.79 ∗ −0.37 ∗ −1.14 ∗ Gemma-4-31B−0.09 ∗ −0.75 ∗ −0.38 ∗ −1.39 ∗ Gemma-4-E4B+0.00−0.69 ∗ −0.59 ∗ −1.29 ∗ Qwen-3.6-27B−0.09 ∗ −0.57 ∗ −0.33 ∗ −1.08 ∗ Qwen-3.6-35B-A3B−0.03−0.49 ∗ −0.23 ∗ −0.96 ∗ Table 12: Overall composite BCT effects by model and strategy, aggregated across responses from 600 questions, 12 languages, and all BCT features. F Overall Action Distribution Table13reports the forced-choice action distribu- tion (Confrontation, Redirection, Disengagement) acrossallstrategies. TheEnglishbaselineisshown for reference; NL shows the largest shift, mov- ingresponsesawayfromConfrontation andtoward Redirection relative to English. Again, signifi- cance is computed using pairwise comparisons to the English baseline. F.1 Action Distribution by Language Table14breaks down the forced-choice action dis- tribution by language group and strategy. F.2 Action Distribution by Topic Table15reports action distributions by conversa- tion topic. We also report the NP–NL differences condi- tioned on topic and language in Figure13. F.3 Action Distribution by LLM Model Table16reports model-level action distributions. Figure 13: NP–NL differences in forced-choice action rates by topic and language, aggregated across responses from all eight models. With less data in each sample, many of the bars appear not significant. Strategy Confrontation Redirection Disengagement English34.152.013.9 TTG35.450.314.3 NL28.8 ∗ 57.6 ∗ 13.5 ∗ OSP30.1 ∗ 54.6 ∗ 15.3 ∗ NP32.353.414.3 Table13: Forcedchoiceactionratesforthefivelanguagepathways, aggregatedacrossresponsesfrom600questions, 12 languages, and eight models. Language group ActionTTG NL OSP NP East AsianConfrontation33.0 23.0 ∗ 25.0 ∗ 28.0 ∗ Redirection51.6 61.2 ∗ 57.5 ∗ 57.0 ∗ Disengagement15.4 15.8 ∗ 17.5 ∗ 15.0 ∗ Mid. East/Turkic Confrontation36.1 27.5 ∗ 28.7 ∗ 31.5 ∗ Redirection50.2 58.7 ∗ 56.2 ∗ 55.3 ∗ Disengagement13.7 13.8 ∗ 15.1 ∗ 13.2 ∗ GermanicConfrontation37.2 ∗ 34.1 ∗ 35.1 37.0 ∗ Redirection49.3 ∗ 54.0 ∗ 51.7 49.6 ∗ Disengagement13.5 ∗ 11.9 ∗ 13.3 13.5 ∗ Romance Confrontation 35 . 3 30 . 7 ∗ 32 . 2 33 . 4 Redirection50.2 56.7 ∗ 52.5 50.8 Disengagement14.5 12.7 ∗ 15.2 15.8 Table 14: Forced choice action rates by language group and strategy, aggregated across responses from 600 ques- tions and eight models. TopicActionTTG NL OSP NP WorkConfrontation28.4 23.8 25.3 24.7 Redirection57.8 62.9 60.2 61.3 Disengagement13.8 13.3 14.5 14.0 FriendsConfrontation37.3 32.2 ∗ 33.2 ∗ 35.4 Redirection44.8 50.6 ∗ 47.3 ∗ 46.6 Disengagement17.9 17.2 ∗ 19.6 ∗ 18.0 Relationships Confrontation44.0 36.6 ∗ 37.6 ∗ 42.0 Redirection42.3 50.6 ∗ 47.9 ∗ 43.6 Disengagement13.7 12.9 ∗ 14.5 ∗ 14.5 FamilyConfrontation32.7 23.6 ∗ 25.3 ∗ 28.1 ∗ Redirection55.5 65.5 ∗ 61.7 ∗ 60.7 ∗ Disengagement11.8 10.8 ∗ 13.0 ∗ 11.2 ∗ Table 15: Forced choice action rates by topic and strategy, aggregated across responses from 12 languages and eight models. ModelActionTTG NL OSP NP Claude Opus-4.6Confrontation38.4 19.3 ∗ 24.3 ∗ 30.4 Redirection52.8 74.3 ∗ 68.4 ∗ 62.2 Disengagement8.8 6.4 ∗ 7.4 ∗ 7.4 GPT-4oConfrontation39.1 31.7 32.7 37.1 Redirection45.9 53.8 53.1 49.6 Disengagement15.0 14.6 14.3 13.3 GPT-5.5Confrontation35.2 29.4 31.1 36.8 Redirection53.1 57.8 54.6 50.9 Disengagement11.7 12.8 14.3 12.3 Gemma-4-26B-A4B Confrontation26.1 25.5 24.5 ∗ 28.4 Redirection57.5 56.7 55.4 ∗ 55.6 Disengagement16.4 17.9 20.1 ∗ 16.0 Gemma-4-31BConfrontation48.9 46.3 45.9 45.4 Redirection38.7 41.7 40.7 42.1 Disengagement12.4 12.0 13.4 12.5 Gemma-4-E4BConfrontation24.1 ∗ 18.7 ∗ 20.2 18.0 Redirection57.9 ∗ 64.8 ∗ 60.8 61.7 Disengagement18.0 ∗ 16.5 ∗ 19.0 20.3 Qwen-3.6-27BConfrontation34.6 29.0 30.1 30.3 Redirection49.5 57.2 53.0 53.6 Disengagement15.9 13.8 16.9 16.1 Qwen-3.6-35B-A3B Confrontation36.8 30.8 32.0 32.0 Redirection47.0 54.8 50.8 51.5 Disengagement16.2 14.3 17.2 16.5 Table 16: Forced choice action rates by model and strategy, aggregated across responses from 600 questions and 12 languages. G Sensitivity to Persona-Prompt Wording We test whether the main NP–NL findings depend on the exact wording of the native-speaker persona. In addition to the original NP formulation, we eval- uate two controlled variants that modify only the native-speaker identity clause: •Original NP:”You are a native Language speaker.” •Grew-up NP:”You are a native Language speaker who grew up using that language.” •Current-use NP:”You are a native Lan- guage speaker who currently uses that lan- guage.” We evaluate the three formulations on a stratified subset of 100 prompts, with 25 prompts per topic, across the 12 non-English languages and two models:claude-opus-4.6 andgoogle/gemma-4-E4B-it. We use the same matched comparisons and significance-testing procedures as in the main analysis. The full NP and NL generation procedures are described earlier in Appendix A.3. Figure 14: NP–NL differences in LLM-annotated lin- guistic features by persona formulations, aggregated across responses from 100 questions, 12 languages, and two models. Figures14and15report the linguistic results. Several effects recur under all three formulations. Relative to NL, each NP formulation is signif- icantly more direct, less formal, and less con- crete. All three also use significantly less power- related language and more cognitive-process lan- guage, tentativeness, and social references. Other effects are more sensitive to wording: reductions in emotional expressiveness and social attunement weaken under Current-use NP, while effects for achievement, certitude, prosocial language, and conflict vary in significance across formulations. These results are consistent with the broader lin- guistic patterns. Figure 16shows that all eight BCT dimensions shift in the same direction under every formula- tion: each NP variant provides less behavioral scaffolding than NL. Six dimensions are signifi- cant under all three formulations, including goal setting, action planning, consequence reasoning, self-monitoring, problemsolving, andcognitivere- framing. Forced-choice effects vary more in magni- tude, but all three formulations remain more confrontation-oriented and less redirection- oriented than NL, as shown in Figure 17. Overall, the principal directional findings persist across the controlled wording changes, although their magnitudes and statistical significance remain prompt-sensitive. Figure 15: NP–NL differences in LIWC linguistic fea- tures by persona formulations, aggregated across re- sponses from 100 questions, 12 languages, and two models. Figure 16: NP–NL differences in BCT features by per- sona formulations, aggregated across responses from 100 questions, 12 languages, and two models. Figure 17: NP–NL differences in forced-choice action rates by persona formulations, aggregated across re- sponses from 100 questions, 12 languages, and two models. H Sensitivity to Translation Quality We further assess whether the observed NP–NL differences could be driven by errors in the trans- lated prompts. We repeat the main analyses af- ter retaining only prompt–language pairs whose se- lected translation has a mean back-translation sim- ilarity of at least 0.95. The filter is applied at the prompt–language level before comparing NP and NL responses and retains 4,015 of 7,200 pairs (55.8 percent). Table 17reports the retained sam- ple by language. Because retention rates differ across languages, directly pooling the filtered observations would change the language composition of the analysis. We therefore estimate NP–NL contrasts separately within each language, using the same matched prompt–model comparisons as in the main analy- sis. Figure 18compares the five LLM-annotated di- mensions before and after filtering. Of the 60 language–dimension contrasts, 59 retain their di- rection; themeanabsolutechangeisapproximately 0.03z-score units, and the largest change is 0.13. The broader pattern remains consistent: NP is gen- erally more direct, while NL is generally more for- mal, emotionally expressive, socially attuned, and concrete. The LIWC results are also directionally stable (Figure 19). Of the 144 language–dimension con- trasts, 131 retain their direction. Twelve of the 13 reversals involve contrasts with estimates no larger than 0.09 in absolute value, indicating fluctuation around near-zero differences rather than substan- tive changes. The more pronounced lexical pat- terns remain: NP generally uses more affiliation, positive tone, social references, and prosocial lan- guage, while using less achievement and negative- Language Retained pairs Retained (%) Japanese13923.2 Korean26143.5 Chinese17829.7 Arabic36460.7 Persian27345.5 Turkish30651.0 German39065.0 Dutch41368.8 Swedish37963.2 Spanish46377.2 Italian41469.0 Portuguese43572.5 Overall4,01555.8 Table 17: Prompt–language pairs retained by the translation-quality filter. Each language has 600 pairs before filtering. Retention requires mean back-translation similarity≥0.95. Figure 18: NP–NL differences in LLM-annotated linguistic features by language and translation quality, aggregated across responses from 600 questions and eight models. tone language than NL. Most instability occurs among dimensions for which the original NP–NL differences are already small. Figure 20reportstheeightBCTdimensions. All 96 language–dimension contrasts remain negative after filtering, indicating less behavioral scaffold- ing under NP than under NL in every language and dimension. The mean absolute change is approxi- mately 0.02z-score units, with a maximum change of 0.09. The forced-choice results preserve the main ag- gregate pattern (Figure 21) as well. At the lan- guage level, the confrontation and redirection con- trasts retain their directions in all 12 languages, whereas the smaller base disengagement contrasts are less stable. Overall, restricting the analysis to high- similarity prompt translations produces little change in the principal NP–NL patterns. Nearly all linguistic contrasts retain their direction, every BCT contrast remains negative, and the shift from redirection toward confrontation persists. Figure 19: NP–NL differences in LIWC linguistic features by language and translation quality, aggregated across responses from 600 questions and eight models. Figure 20: NP–NL differences in BCT features by language and translation quality, aggregated across responses from 600 questions and eight models. Figure 21: NP–NL differences in forced-choice action rates features by language and translation quality, aggregated across responses from 600 questions and eight models. I Cross-LLM-as-a-Judge Comparison WetestwhethertheLLM-annotatedresultsdepend on the choice of automated evaluator. In addi- tion to the originalgpt-4oannotations, we anno- tate the same responses usingclaude-sonnet-5 withidenticallinguistic-styleandBCTrubrics. We evaluate a stratified subset of 100 prompts, with 25 prompts per topic, across the 12 non-English languages and three generation models:gpt-4o, claude-opus-4.6, andgoogle/gemma-4-E4B-it. We use the same matched comparisons and significance-testing procedure as in the main analy- sis. Because LIWC features and forced-choice out- comesdonotrelyonanLLMevaluator, thisrobust- ness check applies only to the five linguistic-style and eight BCT dimensions. Figure22compares the linguistic-style effects estimated by the two judges. At the aggregate level, the judges agree on the direction of all five NP– NL effects. At the language level, they agree on 58 of the 60 language-by-dimension effects (96.7%), and their standardized effect sizes are strongly cor- Figure 22: NP–NL differences in LLM-annotated lin- guistic features by LLM-as-a-judge model, aggregated across responses from 100 questions, 12 languages, and three models. related (r=.89). When separated by generation model, the judges agree on 13 of the 15 model-by- dimension effects, with similarly correlated effect sizes (r=.90). Thus, the main linguistic differ- ences are largely stable across evaluators. Figure 23reports the BCT results. The judges agree on the direction of seven of the eight aggre- gate NP–NL effects. These seven dimensions in- dicate less behavioral scaffolding under NP than under NL and are significant under both evalua- tors. The only discrepancy is feedback:gpt-4o estimates a moderate NP–NL reduction, whereas claude-sonnet-5estimates a near-null effect. Agreement is also strong when results are sep- arated by generation model. Critically, agree- ment does not weaken for responses generated by gpt-4o, despitegpt-4oserving as the original evaluator (Figure 24). For these responses, the judges agree on 12 of the 13 effects (r=.77). The only directional difference is directness, for which the original evaluator estimates a near-null effectwhileclaude-sonnet-5estimatesapositive effect. Overall, the principal directional findings per- sist across the two automated evaluators, includ- ing whengpt-4oevaluates its own generated re- sponses. Figure 23: NP–NL differences in BCT features by LLM-as-a-judge model, aggregated across responses from 100 questions, 12 languages, and three models. Figure 24: NP–NL differences in LLM-annotated (lin- guistic and BCT) features on gpt-4o generated re- sponses, aggregated across responses from 100 ques- tions and 12 languages.