Paper deep dive
Wiki-Talkie: Multilingual Benchmarking of Persona-Based Agents on Real-World Discussions
Dennis Fucci, Andrea Bacciu, Dong Liu, Weronika Åajewska, Saab Mansour
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 deployed as autonomous agents in social environments, making it critical to study their ability to faithfully simulate human interactions. Central to this is grounding agents in realistic user personas, yet existing datasets rely on fictional personas and are limited to a handful of languages, lacking the empirical grounding necessary to evaluate behavioral fidelity across diverse populations. We introduce Wiki-Talkie, a multilingual dataset of real-world conversations from Wikipedia Talk pages across five languages spanning two language families: Germanic (German, English) and Romance (Spanish, French, Italian), paired with personas derived from real user communities and encompassing sociodemographic attributes, self-descriptions, and behaviorally grounded interaction traits. Using Wiki-Talkie, we evaluate agent interactional behavior on a next-turn generation task across various persona conditioning strategies. Our evaluation assesses whether agents collectively reproduce the distributional behavioral patterns observed in human discussions. Results show that user's comment history exemplifying interaction behavior consistently outperforms explicit persona information. In addition, models systematically underproduce negative or extreme sentiments, while over producing references and suggestions, revealing biases toward agreeableness and positivity. Crucially, these patterns hold robustly across languages, with small cross-lingual differences.
Tags
Links
- Source: https://arxiv.org/abs/2610.08513v1
- Canonical: https://arxiv.org/abs/2610.08513v1
Trouble viewing inline? Open PDF directly ā
Full Text
145,815 characters extracted from source content.
Expand or collapse full text
Wiki-Talkie: Multilingual Benchmarking of Persona-Based Agents on Real-World Discussions Dennis Fucci ā thanks: Work done during an internship at Amazon. Email: dennisfucci@gmail.com Affiliation: Andrea Bacciu Dong Liu Weronika Åajewska Saab Mansour Affiliation: Amazon Affiliation: andbac, liuadong, lajewska, saabm@amazon.com Abstract LLMs are increasingly deployed as autonomous agents in social environments, making it critical to study their ability to faithfully simulate human interactions. Central to this is grounding agents in realistic user personas, yet existing datasets rely on fictional personas and are limited to a handful of languages, lacking the empirical grounding necessary to evaluate behavioral fidelity across diverse populations. We introduce Wiki-Talkie, a multilingual dataset of real-world conversations from Wikipedia Talk pages across five languages spanning two language families: Germanic (German, English) and Romance (Spanish, French, Italian), paired with personas derived from real user communities and encompassing sociodemographic attributes, self-descriptions, and behaviorally grounded interaction traits. Using Wiki-Talkie, we evaluate agent interactional behavior on a next-turn generation task across various persona conditioning strategies. Our evaluation assesses whether agents collectively reproduce the distributional behavioral patterns observed in human discussions. Results show that userās comment history exemplifying interaction behavior consistently outperforms explicit persona information. In addition, models systematically underproduce negative or extreme sentiments, while over producing references and suggestions, revealing biases toward agreeableness and positivity. Crucially, these patterns hold robustly across languages, with small cross-lingual differences.11 1 Code and data are available at https://github.com/amazon-science/wikitalkie. 1 Introduction Large Language Models (LLMs) are increasingly embedded in real-world social systems, from community moderation on Reddit22 2 https://redditinc.com/news/oterlu-team-joins-reddit-to-accelerate-safety-efforts. to experimental platforms like MoltBook33 3 https://w.moltbook.com. where LLMs autonomously interact at scale. These developments position LLMs not merely as tools, but as agents: autonomous actors participating in governance, culture, and social interaction. Central to the design of such agents is the role they are asked to play (Shanahan et al., 2023; Tseng et al., 2024), e.g. the persona, that is, a composite of individual traits such as background facts, language behavior, and interaction style (Liu et al., 2025a). This approach has been shown to improve alignment with real human populations in applications such as personalized dialogues (Liu et al., 2025a), linguistic annotation (Orlikowski et al., 2025), and social simulations (Argyle et al., 2023; Park et al., 2024). However, existing datasets for modeling persona-based conversational agents rely on fictional personas (Smith et al., 2020; Jang et al., 2022; Xu et al., 2022a; Xu et al., 2022b; Yamashita et al., 2023; Zhao et al., 2025; Lotfi et al., 2023; Chen et al., 2022; Occhipinti et al., 2024), making it impossible to verify whether agent behavior faithfully reflects the opinions, linguistic patterns, and interaction dynamics of the actual human population being simulated. Datasets extracting personas from large-scale media (Zheng et al., 2020; Wu et al., 2021) offer greater realism but are limited to Chinese or English. To our best knowledge, there is no resource providing real-world, multilingual dialogues grounded in real personas within a cooperative setting where pragmatic phenomena such as conflict, misalignment, and evolving viewpoints can emergeāprecisely the dynamics that characterize high-stakes scenarios. To fill this gap, we introduce Wiki-Talkie, a multilingual dataset of real-world discussions extracted from Wikipedia Talk pages across five languages, where registered users collaboratively deliberate on article edits. Wiki-Talkie enables multifaceted evaluation of agents in a consensus-oriented discussion scenario, with optional conditioning on personas derived from Wikipedia user pages, and behavioral traits from comment histories. To safeguard user privacy, all identities are anonymized and persona profiles are regenerated in a neutral fashion rather than verbatim copies, ensuring privacy protection while maintaining research utility. Our evaluation targets what we term population fidelity: whether agents collectively reproduce the distributional behavioral patterns observed in human discussions, as opposed to individual-level persona fidelity, i.e., whether a single generated response faithfully reflects the provided persona of a specific user. This choice is principled: individual-level fidelity risks enabling faithful simulation of specific individuals, raising ethical concerns around impersonation and lack of consent (Desai et al., 2025), while population fidelity is more tractable and more relevant for applications such as community simulation and policy analysis. This aligns with an emerging consensus that evaluation of persona-based agents should operate at the population level (Mansour et al., 2025; Rupprecht et al., 2025; Fawaz et al., 2026). To assess population fidelity, we frame the task as next-turn generation: given a thread of previous comments, an agent must generate the next comment simulating a specific Wikipedia editor. We conduct evaluations with multiple metrics: instance-level text similarity (ROUGE-L and BERTScore), which provides individual-level signals averaged across the entire dataset, and population-level interactional behavior, comparing sentiment and dialog act distributions (Searle, 1969) between human comments and those generated by agents powered by multiple state-of-the-art LLMs, across five languages. We study how agents behave when conditioned on different types of persona information, comparing behavioral traits extracted from comment histories against explicit persona descriptions derived from user profiles. Results show that cross-lingual differences are small, consistent with prior work on multilingual LLM evaluation showing strong performance across high-resource languages Xuan et al. (2025); Li et al. (2025). We also find that comment history consistently proves more effective than explicit persona information as a conditioning strategy, suggesting that showing how a persona behaves outperforms telling what a persona is. Overall, our contributions are: 1) the construction of Wiki-Talkie, a dataset of Wikipedia discussions supporting research on agentic interaction simulation, whose key advantages are: a) multilingual coverage; b) real-world discussions reflecting established communication norms; c) ground-truth persona information linked to discussions; and d) rich contextual information, with discussions paired with articles; 2) the evaluation of agents based on four LLMs under various conditioning strategies across five languages, providing insights into the limitations of current agentic simulation of human cooperative interactions. 2 Related Work Datasets for Persona-based Evaluation Conversational datasets with persona-related attributes vary in source, language, and use case. Crowdsourced datasets dominate, exemplified by PersonaChat Zhang et al. (2018b), where crowdworkers converse under fictional personas, and its many extensions: BlendedSkillTalk (Smith et al., 2020) adds empathy and knowledge, FoCus (Jang et al., 2022) adds Wikipedia grounding, Multi-Session Chat (Xu et al., 2022a) models longer-term interaction, DuLeMon (Xu et al., 2022b) uses Chinese mutual personas, and RealPersonaChat (Yamashita et al., 2023) collects real persona descriptions in Japanese. Other datasets rely on synthetic generation (Zhao et al., 2025; Lotfi et al., 2023) or scripted content (Chen et al., 2022; Occhipinti et al., 2024), while social media datasets such as PersonalDialog (Zheng et al., 2020) and PER-CHAT (Wu et al., 2021) extract persona signals from large-scale platforms. Most rely on simulated role-play or synthetic generation and cover few languages: although XPersona (Lin et al., 2021) extends PersonaChat multilingually via automatic translation, naturally occurring interaction datasets remain largely restricted to English and Chinese. Critically, no existing resource simultaneously provides: (i) naturally occurring dialogues, (i) multilingual coverage, (i) real user-derived personas linked to the conversations, and (iv) a cooperative yet contentious setting where disagreement, negotiation, and consensus-building emerge. We address this gap with Wiki-Talkie. LLM Social Behavior A growing body of work documents systematic biases in LLM-generated social behavior. LLMs exhibit sycophantic tendencies, preferring agreement over truthful disagreement (Malmqvist, 2025), and display a positivity bias that suppresses negative or confrontational outputs (Santurkar et al., 2023). Studies using LLMs to simulate human populations have shown that, while persona conditioning can improve alignment with survey response distributions (Argyle et al., 2023), models often fail to reproduce the full diversity of human opinions and behaviors (Orlikowski et al., 2025). However, these findings are predominantly based on English-only settings, synthetic survey tasks, or opinion elicitation, leaving open the question of whether such biases persist in naturalistic, multilingual, cooperative discussions. Wiki-Talkie addresses this gap by providing a testbed where LLM social behavior can be evaluated against real human interactions across five languages in a collaborative setting with genuine disagreement and pragmatic complexity. Datasets Derived from Wikipedia Talks Wikipedia Talk pages remain underexplored despite providing large-scale, naturally occurring collaborative dialogue Johnson et al. (2024). Prior work has used them to study conversational behavior and harmful interactions (Wulczyn et al., 2017; Zhang et al., 2018a), dialogue act classification Ferschke et al. (2012), dispute detection Wang and Cardie (2014), sociolinguistics Prabhakaran and Rambow (2016), abuse (CĆ©cillon et al., 2020), and links between discussions and edit histories (Jaidka et al., 2021). WikiConv (Hua et al., 2018) reconstructs full conversational histories including edits and deletions, supporting tasks such as preemptive toxicity detection (Karan and Å najder, 2019). However, these datasets focus on discourse structure, conflict, or toxicity, remain mostly English-centric, and lack persistent persona representations for multilingual evaluation. To our knowledge, Talk pages have not been used to analyze LLM behavior in socially grounded, collaborative settings. 3 Wiki-Talkie Figure 1: Example from Wiki-Talkie with its three core components: article, thread, and persona. We introduce Wiki-Talkie, a multilingual dataset of real-world discussions extracted from Wikipedia Talk pages across five languages spanning two language families: Germanic (German, English) and Romance (Spanish, French, Italian). Wiki-Talkie is built around three core components (see Figure 1): 1) multilingual threads in which users deliberate on article content; 2) user personas from two complementary sources, a self-reported profile derived from Wikipedia user pages (with attributes such as gender, profession, interests, and contribution topics) and behavioral traits inferred from each userās comment history; and 3) the article context, paired with each thread as it appeared immediately before the discussion began. Together, these components make Wiki-Talkie suited for research on persona-conditioned and document-grounded interaction modeling across languages.44 4 Examples are in Appendix B, Tables 6ā10. Table 1 presents summary statistics of Wiki-Talkie across languages (full distributions in Appendix A.1.1). German contributes the largest portion (>42K threads, >134K turns), followed by French and English, while Spanish and Italian form smaller subsets. Despite differences in scale, structural characteristics are remarkably consistent across languages: threads contain on average 3.2--3.6 turns and involve 2.1--2.3 unique users.55 5 These statistics are not necessarily representative of Wikipedia Talk pages at large, but reflect our construction pipeline (Section 3.1). Turn length varies more, ranging from 47.74 words (German) to 79.73 (Spanish), potentially reflecting differences in editorial culture or linguistic verbosity. High standard deviations across all metrics indicate substantial variability, consistent with discussions ranging from brief clarifications to extended debates. The number of unique users mirrors the thread distribution; further details on users and personas are in Appendix A.1.2. # Threads # Turns # Users Turns/Thread Turns/User Turn Length Users/Thread de 38,711 123,520 6,282 3.19 Ā± 1.92 19.93 Ā± 53.93 48.09 Ā± 56.80 2.11 Ā± 0.35 en 17,645 59,537 4,580 3.37 Ā± 2.33 13.19 Ā± 40.01 67.30 Ā± 81.68 2.06 Ā± 0.24 es 5,229 17,019 2,307 3.25 Ā± 2.13 7.55 Ā± 19.41 79.80 Ā± 96.36 2.14 Ā± 0.42 fr 23,983 86,041 4,952 3.59 Ā± 2.37 17.59 Ā± 58.68 66.32 Ā± 79.56 2.26 Ā± 0.61 it 5,453 18,956 1,850 3.48 Ā± 2.35 10.39 Ā± 27.04 66.00 Ā± 84.43 2.18 Ā± 0.46 Table 1: Summary statistics of Wiki-Talkie ground-truth threads, turns, users, and average metrics (mean Ā± std). 3.1 Dataset Construction Threads The threads contained in Wiki-Talkie are extracted from raw Wikipedia Talk page dumps.66 6 Dumps from October 2025: https://dumps.wikimedia.org/<lang_code>wiki/20251001/. Wikipedia Talk pages are largely unstructured, making clean conversation extraction a challenging preprocessing task Ferschke et al. (2012); Hua et al. (2018). We exploit standard conventions: MediaWiki heading syntax for topics, username-timestamp signatures, and colon-based reply indentation, and handcrafted regular expressions to parse threads into flat sequential turns ordered by timestamps 77 7 See https://en.wikipedia.org/wiki/Wikipedia:Talk_page_guidelines.. To address extraction errors, we prompted the extractor model to revise threads by correcting turn boundaries and comment attributions, flagging irrecoverable cases for exclusion (prompt reported in Appendix C, Figure 10). Persona Profiles For each unique user in the extracted threads, we fetched the corresponding Wikipedia user page via the MediaWiki API.88 8 https://w.mediawiki.org/wiki/API, accessed through mwclient (https://github.com/mwclient/mwclient). User pages are optional personal spaces where editors describe themselves, their interests, expertise, and contributions. We prompted the extractor model to determine whether a page contained relevant biographical or self-descriptive content, excluding pages consisting solely of administrative templates or userboxes (prompt in Appendix C, Figure 15). For pages with substantive content, the model also generated a natural-language profile summarizing all available details, following recent work on LLM-based persona extraction (Zhang, 2024; Wu et al., 2025; Mansour et al., 2025). Because these are self-declared attributes from real users, the resulting personas avoid the stereotypical or synthetic quality of artificially constructed profiles and naturally incorporate sociodemographic aspects, making them valuable for evaluating how such factors influence agent behavior and biases. Alongside the profile, the model extracted predefined attributes in JSON format (e.g., primary editing topics, gender, profession) to support statistical analysis of user subgroups. Users with successfully generated profiles were subsampled to at most 10,000 per language, prioritizing presence of contribution topics, then gender (favoring women and non-binary users to counteract imbalance), then profession. Only threads where all participants belong to the selected subset were retained. Persona Behavioral Traits From each userās comments (across up to 50 randomly sampled threads), we extracted behavioral traits using the extractor model, characterizing each user along eight dimensions including (dis)agreement strategy and editing behavior. Each dimension is filled with closed response options only when sufficient evidence is available, accompanied by the modelās reasoning (prompts in Appendix C, Figures 11ā13). Unlike self-declared profiles, these traits are inferred from concrete interactions, making them especially useful for evaluating how behavioral conditioning shapes agent output. A manual quality audit of profiles and traits (faithfulness, coherence, privacy preservation) is reported in Appendix I. Articles For each thread, the discussed Wikipedia article is retrieved via the MediaWiki API to provide conversational context, using the revision immediately preceding the first commentās timestamp. This temporal alignment ensures the retrieved content reflects the article state when the discussion began. Finally, all usernames are anonymized via deterministic hashing to prevent re-identification, and thread and article formatting is cleaned and normalized for consistency. Full details are in Appendix A.2. 4 Experimental Setting Task Wiki-Talkie supports tasks including comment generation, summarization, and quality assessment. Here, we frame the task as next-turn generation: given the preceding comments in a thread, an agent is asked to generate the next comment in the role of the real user who authored it. We always condition on at least one preceding comment (i.e., the first turn is never generated from scratch), and prior context always consists of ground-truth comments. This task formulation was chosen deliberately to enable controlled, comparable evaluation across languages and conditioning strategies, resulting in over 100 experimental conditions (5 languages Ć 4 models Ć 5 conditioning methods). Although Wiki-Talkie supports evaluation along several dimensionsāincluding coherence with context, factual grounding in the associated article, and fidelity to persona conditioningāhere we focus on interaction behavior of agents. Data To keep generation costs manageable, we experimented with a subsample of Wiki-Talkie. We selected threads involving users with at least six comments in the corpus, ensuring a minimum five-comment history prior to the target comment. From these, we sampled up to 5,000 threads per language, maximizing the proportion of comments authored by females up to 50%. Summary statistics are presented in Appendix D. Figure 2: ROUGE-L and rescaled BERTScore between humans and agents across the five conditioning methods. Left: averages across models for each language. Right: averages across languages for each model. Conditioning Methods We defined five conditioning methods to investigate how different types of persona information shape agent behavior in interactions: i) base role-play, where the agent is instructed that it is a Wikipedia editor with no additional information; i) profile, where the agent is provided with the self-reported profile (gender, age, profession, etc.) of the user who produced the target comment; i) behavior, where the agent is provided with the userās behavioral traits; iv) behavior+r, same as behavior but including the extractor modelās reasoning; v) history, where the agent is provided with 5-20 comments by the same user (randomly sampled if more are available), enabling indirect conditioning through stylistic and behavioral patterns in prior contributions. The prompts for each method are provided in Appendix C, Figures 14, 15, 16, 17. Models We run experiments with four state-of-the-art models spanning both open-source and proprietary systems: Model A (proprietary), Model B, Model C, and Model D (open-weights).99 9 Model identities are withheld; we report results under anonymized labels (Model AāModel D). All four are widely-used frontier LLMs at the top of their respective open-weights and proprietary classes at the time of writing, chosen to ensure that our findings reflect the behavior of strong, competitive systems rather than artifacts of weak models. Implementation and generation details are provided in Appendix E. Evaluation Metrics We report text similarity via ROUGE-L (Lin, 2004) for surface lexical overlap and BERTScore (Zhang et al., 2020) for semantic similarity, both raw and rescaled against a baseline (rBERT) as recommended.1010 10 https://github.com/Tiiiger/bert_score. Since generation is open-ended, these are coarse reference points rather than primary criteria. Our primary focus is interactional behavior. We extract dialogue acts using the taxonomy of Ferschke et al. (2012), developed for Wikipedia Talk pages, considering 10 acts: critique, request, reference, commit, report, inform, ask, positive, negative, and partial.1111 11 Mapping to Ferschke et al. (2012) is in Appendix F. Both human and agent comments are annotated with the extractor model as LLM annotator (prompt in Appendix C, Figure 12), which achieves 0.76 F1 against human judgment (Appendix F). Each comment can carry one or multiple acts. We additionally assess overall response tone via sentiment prediction on both human and agent comments.1212 12 Using the multilingual model https://huggingface.co/tabularisai/multilingual-sentiment-analysis, predicting on a five-point scale from very negative to very positive. All metrics are computed at the dataset level. BERTScore and ROUGE-L are averaged over all output-reference pairs. Sentiment and dialogue act distributions are compared via Wasserstein Distance (WD; Panaretos and Zemel, 2019) and Jensen-Shannon Divergence (JSD; MenĆ©ndez et al., 1997), reflecting the ordinal nature of sentiment and the nominal nature of acts. This aggregate perspective targets population fidelityāwhether agents collectively reproduce distributional behavioral tendenciesārather than individual-level persona fidelity. ROUGE-L and BERTScore are computed at the instance level against each target userās ground-truth response and then averaged across the dataset. 5 Results 5.1 Text Similarity Figure 2 shows ROUGE-L and rBERT scores averaged across models and languages; full results, including average comment length and raw BERTScore (which remains stable across setups, ā¼0.84 0.84ā0.850.85, and is thus less informative), are in Appendix G, Tables 11ā12. While our primary target is population fidelity (Section 4), these instance-level metrics provide a complementary individual-level signal. ROUGE-L scores never exceed 0.1540.154. While low in absolute terms, this is expected given the open-ended nature of the task, where the space of plausible responses is virtually unbounded; scores are thus best read as relative measures. rBERT hovers around zero, indicating that agent-human similarity is broadly comparable to random sentence-human pairs, further confirming the difficulty of the task. Figure 3: JSD between human and agent dialog act distributions across conditioning methods. Left: averages across models for each language. Right: averages across languages for each model. Across both languages and models, history consistently yields the highest similarity, with ROUGE-L up to 0.1540.154 (Model A in es) and rBERT up to 0.1050.105 (Model A in de). behavior+r is the second most effective, with consistent gains over base (e.g., ROUGE-L up to 0.1380.138 for Model A in es; rBERT up to 0.0850.085 for Model A in de). By contrast, profile generally degrades similarity relative to base (e.g., rBERT drops on average from 0.0530.053 to 0.0400.040 for Model A, and from ā0.017-0.017 to ā0.049-0.049 for Model D), likely because prompting agents to reflect persona traits leads them to incorporate non-relevant details, producing longer and more elaborated responses without improving content alignment. Indeed, agent comments are consistently longer than human comments across all conditions (Appendix G). behavior similarly fails to improve similarity for most models and languages, with only modest gains in some cases (e.g., Model Bās ROUGE-L from 0.0780.078 to 0.0820.082). Cross-lingually, ROUGE-L scores are consistently higher for es and en than for de and it, partially mirroring human comment length: longer reference texts provide richer surface for lexical overlap. rBERT trends are less pronounced; en falls slightly below the others, largely driven by Model Bās particularly low English scores. Overall, concrete demonstrationsāthrough comment histories (history) or behavioral traits paired with reasoning (behavior+r)āare more effective than abstract descriptions (behavior) and sociodemographic information (profile) at steering agent output toward human-like text. This is the dominant pattern across models and languages, while cross-lingual differences remain small, though we note that all five languages are high-resource and European. The consistency is nonetheless non-trivial given that they span two language families and differ in Wikipedia community size and editorial norms. 5.2 Interaction Behavior Figure 4: Proportion difference between human and agent dialog acts. Positive values indicate human over-production; negative values indicate agent over-production. Differences are averaged across languages. Figure 5: Distribution of sentiment scores for human and agent (Model A under history) comments for en. In the following we discuss evaluation of behavioral patterns through dialog act JSD and sentiment WD; detailed scores are reported in Appendix G, Tables 13 and 14. Acts JSD. Figure 3 presents average JSD scores aggregated across models (left) and languages (right). Scores range between 0.10.1 and 0.30.3, indicating only moderate divergence between human and agent dialogue act distributions. history is the strongest method, achieving the lowest JSD overall (0.0850.085 for Model A in en). behavior+r is second, with consistent gains over base (e.g., Model B: 0.272ā0.2410.272ā 0.241; Model D: 0.199ā0.1660.199ā 0.166). By contrast, behavior is the weakest, increasing JSD across most models and languages (except Model B, where it improves over base, 0.272ā0.2450.272ā 0.245). profile yields only marginal effects, with gains limited to Model A (0.150ā0.1470.150ā 0.147). These trends are broadly consistent across languages, though en achieves the lowest average JSD throughout (e.g., 0.1880.188 under base, 0.1750.175 under behavior+r), likely reflecting English-centric pretraining. Cross-lingual differences, however, remain marginal. Overall, these results reinforce the pattern observed for text similarity: concrete behavioral demonstrations (history, behavior+r) are consistently more effective than abstract descriptions or sociodemographic factors (behavior, profile) at aligning agent behavior with human distributions, echoing the well-established advantage of few-shot over zero-shot conditioning Brown et al. (2020). One might wonder whether this gap reflects a difference in information type or simply in information quantity, since history provides 5ā20 raw comments while profile and behavior supply compressed summaries. We address this concern in Appendix H, where we show that performance under profile does not improve consistently with profile length: ROUGE-L and rBERT are flat, and where acts and sentiment JSD change, they mostly worsen with longer profiles. Combined with the observation that profile degrades performance relative to base despite adding tokens, this argues against a pure context-budget explanation and supports our interpretation that the operative factor is information type, not quantity. Per-act differences To assess whether divergences concentrate on specific labels, we analyze per-label proportion differences (Ī“) between human and agent distributions, where positive values indicate human over-production and negative values indicate agent over-production. Figure 4 shows differences averaged across languages, omitting acts with negligible variation (Ask, Critique, Positive, Partial). Differences are computed between proportions rather than raw counts to account for lower act density in agent comments (Appendix G). Full breakdowns are in Tables 13ā14. Agents systematically over-produce Request (up to Ī“=ā0.177Ī“=-0.177 for Model C in de on behavior) and Commit (up to Ī“=ā0.108Ī“=-0.108 for Model D in fr on behavior), while humans produce more Report (Ī“=0.088Ī“=0.088 for Model C in fr on behavior). Together, this suggests that agents tend to operate by requesting and committing to changes, yet report fewer actions already undertaken. Inform is under-produced by agents across most models, possibly reflecting insufficient factual grounding (Model A and Model D diverge in some methods). Reference shows a model-dependent pattern. Negative is consistently under-produced (Ī“ from 0.0200.020 to 0.1040.104), consistent with sycophantic tendencies of LLMs Malmqvist (2025). Overall, the most systematic biases concern operativity (over-production of Request, Commit) and negativity (under-production of Negative, Report), pointing to a consistent tendency for agents to be more action-oriented and less confrontational than humans. Per-act differences by gender To examine whether agents reproduce dialog act trends equitably across genders, we analyze per-label proportion differences split by gender for two models with contrasting alignment levels (Model A, higher; Model B, lower) and two languages with similar contrast (English, higher; French, lower), under history (best overall), profile (the only method with explicit gender), and base as reference. Differences are shown in Appendix G, Figure 18. Trends are consistent across genders for all labels and methods. Crucially, profile does not behave systematically differently from base or history, indicating that explicit gender conditioning neither amplifies nor reduces gender disparities in dialog act profiles. Sentiment Sentiment alignment broadly mirrors the dialog act trends: history yields the best alignment (minimum WD 0.0970.097 for Model C in en), followed by behavior+r (0.1680.168 in en). Across languages, en shows the lowest WD (0.3150.315 on average under base) and es the highest (0.4520.452), consistently across methods. As shown in Figure 5 for Model A under history, agent responses rarely exhibit sentiment extremes (very positive or very negative) and contain fewer negative comments than humans, consistent with their lower rate of Negative acts. These patterns hold across models, conditioning methods, and languages (Appendix G, Figure 19). Overall, sentiment converges with the broader finding that agents tend to avoid negative or confrontational content across all conditioning methods. 6 Conclusions We introduced Wiki-Talkie, a multilingual dataset derived from Wikipedia Talk pages to evaluate agent interactions across five languages, using four state-of-the-art LLMs under different persona conditioning strategies. Our evaluation targets population fidelityāwhether agents collectively reproduce the distributional behavioral patterns of human communitiesācomplemented by instance-level text similarity metrics. Our experiments show that grounding agents with concrete comment examples consistently outperforms abstract persona descriptions across all languages. Notably, sociodemographic information can degrade agent performance, an encouraging finding for privacy-preserving applications. In addition, agents are consistently more action-oriented and less confrontational than humans, revealing a bias toward agreeableness and avoidance of extremes. Cross-lingual differences remain small, with English showing only slightly greater alignment. Overall, these results underscore both the promise and limitations of persona-conditioned generation, motivating further research into more effective simulation strategies. While we focused on next-turn generation to enable controlled comparison across a large experimental space, Wiki-Talkie is designed to support richer evaluation settings, including multi-turn generation, summarization, factual grounding assessment, and quality evaluation, all of which we encourage the community to explore. Limitations Dataset Construction Despite our multi-stage extraction and revision pipeline, some residual noise remains in Wiki-Talkie, including occasional misattributed comments, imperfect comment boundaries, and formatting artifacts not fully resolved during preprocessing Such noise is inevitable at this scale and is unlikely to substantially affect evaluation or analysis. Language Coverage Wiki-Talkie currently covers five European languages (English, German, Spanish, French, and Italian), limiting its cultural and linguistic diversity. While these languages span two distinct families (Romance and Germanic) and vary in Wikipedia community size and editorial norms, they are all high-resource; findings may not generalize to language communities with different writing systems, discourse conventions, or editorial practices. Future work should extend Wiki-Talkie to additional languages, particularly those underrepresented in NLP research. The dataset construction pipeline and evaluation framework are fully language-agnostic and straightforwardly extensible to lower-resource and non-European languages. Dataset Richness The richness and consistency of user personas vary considerably within and across languages. Some profiles contain detailed self-reported attributes, while others are sparse. English is generally better represented in both quantity and quality, meaning persona conditioning experiments may be more reliable for English than for other languages. Moreover, despite subsampling to increase the representation of women and non-binary users, the dataset remains gender-skewed, which may affect the robustness of gender-based analyses. Persona Conditioning Our experiments focus exclusively on prompting-based persona conditioning, leaving alternative approaches to future work, such as fine-tuning (e.g., Thakur et al., 2024; Takayama et al., 2025), retrieval-augmented conditioning (e.g., Huang et al., 2023; Salemi et al., 2024), or embedding-based persona representations (e.g, Liu et al., 2025b). Information Density across Conditioning Methods Our conditioning methods are not strictly matched on information density: history provides 5ā20 raw comments, while profile and behavior supply compressed summaries. We probe this confound in Appendix H by stratifying profile performance by profile length and find no consistent improvement with profile length across four metrics, supporting an information-type rather than information-quantity interpretation. We did not perform the analogous analysis for history length because the conditioning pool is concentrated in the 16ā20 comment range, leaving lower-comment buckets too sparse to support a meaningful comparison. A fully matched ablationāe.g., comparing history against equally long synthetic summaries distilled from the same commentsāremains a direction for future work. Scope of Persona Representations Our profiles and behavioral traits are produced by a single anonymized extractor model and a specific prompting design. A manual audit (Appendix I) indicates that these representations are largely faithful and well-grounded, with imprecisions limited to isolated factual details. Nonetheless, our findings about profile and behavior should be read as conclusions about sociodemographic and behavioral descriptors as we generated them, rather than as universal claims about all possible persona representations. Alternative summarization strategies, richer attribute schemas, or different extractor models could in principle yield different results. Evaluation Scope Our primary evaluation target is population fidelityāwhether agents collectively reproduce the distributional behavioral patterns observed in human discussionsārather than individual-level persona fidelity, i.e., whether each generated response faithfully reflects the specific persona provided. This choice is principled: individual-level fidelity risks enabling faithful simulation of specific individuals, raising ethical concerns around impersonation and lack of consent (Desai et al., 2025), while population fidelity is more tractable and more relevant for applications such as community simulation and policy analysis, in line with recent work advocating population-level evaluation of persona-based agents (Mansour et al., 2025; Fawaz et al., 2026; Rupprecht et al., 2025). Assessing fine-grained individual-level persona consistency remains a direction for future work. Ethic Statement Privacy, Data Protection, and Contextual Integrity Wiki-Talkie is derived from Wikipedia dumps, released under a Creative Commons Attribution-ShareAlike 4.0 license. While publicly accessible, we recognize that contributors may reasonably expect their comments to be archived for research purposes but not necessarily used for persona-conditioned simulation. Accordingly, we treat this dataset as human-derived data requiring safeguards beyond copyright compliance, guided by the principles of data minimization, proportionality, contextual integrity, and harm reduction. Dataset Construction and Anonymization The dataset does not contain original user pages verbatim. Instead, we generated neutral, natural-language re-elaborations summarizing publicly stated interests and interactional patterns, deliberately omitting usernames, real names, contact information, and external links. This process inherently introduces a layer of abstraction: the resulting profiles are approximate characterizations that may diverge from the original content through omission, rephrasing, or occasional inaccuracy. While this means profiles should not be treated as ground-truth descriptions of actual persons, we acknowledge that the same process may occasionally introduce erroneous attributes, including sensitive ones. We sought to mitigate this risk through manual inspection of a random sample of outputs (see Appendix I). Usernames are replaced using salted cryptographic hashing, where the salt is stored separately from the dataset and is accessible only to the principal investigator, preventing straightforward reversal or precomputation attacks (see Appendix A.2). Within comment text, we also anonymized mentions of third-party usernames to the extent possible. However, a residual risk remains where usernames appear in non-standard forms or where real names are mentioned in conversation. Re-Identification Risk and Access Control Complete anonymity cannot be guaranteed when working with public textual data. While salted hashing prevents trivial username reconstruction, a content-based re-identification risk exists: because Wikipedia discussions are publicly archived and indexed, motivated individuals could potentially identify original authors via distinctive phrases or search queries. Our dataset does not introduce new exposure beyond what is already publicly accessible, but we recognize that aggregation and structuring may lower the cost of such searches. To mitigate these risks, the dataset is released under the C BY-SA 4.0 license, inherited from the underlying Wikipedia-derived content. Rather than hosting an open public mirror, we distribute it through a lightweight access request in which prospective users identify themselves and their intended use. This provides an audit trail and a point of contact while raising the effort required for indiscriminate bulk access. As a matter of research norms rather than a binding condition, we ask that recipients not attempt to de-anonymize contributors, not use the data to impersonate or manipulate editorial processes, and work from the most recent release. We provide a dedicated contact address for individuals who wish to request removal of their data. Upon verification, corresponding entries are added to a blocklist and removed in subsequent releases. Our goal is risk mitigation rather than elimination, and we consider the residual risk proportionate to the scientific value of studying collaborative online deliberation at scale. Informed Consent Individual informed consent was not sought, as contributions to Wikipedia are made under the C BY-SA license and the Wikimedia Foundationās Terms of Use, which explicitly permit reuse and derivative works. This is consistent with established practice in NLP research on Wikipedia-derived data (Ferschke et al., 2012; Hua et al., 2018), and the scale of contributors makes individual consent infeasible. Nevertheless, we recognize that licensing permissibility does not fully resolve the ethical tension between contributorsā original communicative intent and downstream research applications; we address this gap through the layered safeguards described above. Harmful or Sensitive Content Wikipedia Talk pages may occasionally contain harassment, conflictual exchanges, or incidental personal disclosures. We do not seek to amplify or foreground such content; the dataset is designed to analyze aggregate interaction patterns in collaborative environments, not for reputational analysis of specific contributors. We did not apply active content filtering in order to preserve the ecological validity of the discussions. Access restrictions are intended to further prevent misuse or amplification of harmful material. Experimental Ethics and Persona Conditioning The objective of this research is not to recreate, approximate, or impersonate specific individuals. Persona-conditioned agents are stochastic constructs derived from coarse behavioral summaries, used solely to examine population-level dynamics in simulated group discussions. Although individual-level conditioning is methodologically necessary to capture heterogeneity in interaction styles, all evaluation and reporting are conducted at the aggregate level. We do not evaluate model fidelity to specific users, publish individual outputs, or claim to model the beliefs, intentions, or identities of real contributors. We recognize that repurposing public discourse for AI research raises novel ethical questions concerning digital representation and contextual integrity. Our ethical justification rests on proportionality: the scientific value of understanding large-scale collaborative behavior, combined with layered safeguards (anonymization, profile re-elaboration, controlled access, opt-out mechanisms), outweighs the residual and mitigated risks. References Argyle et al. (2023) L. P. Argyle, E. C. Busby, N. Fulda, J. R. Gubler, C. Rytting, and D. Wingate Out of one, many: using language models to simulate human samples. Political Analysis 31 (3), p. 337ā351. External Links: ISSN 1476-4989, Link, Document Cited by: Ā§1, Ā§2. Brown et al. (2020) T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Amodei Language models are few-shot learners. In Advances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin (Eds.), Vol. 33, p. 1877ā1901. External Links: Link Cited by: Ā§5.2. Cabrera et al. (2018) B. Cabrera, B. Ross, M. Dado, and M. Heisel The gender gap in wikipedia talk pages. Proceedings of the International AAAI Conference on Web and Social Media 12 (1). External Links: Link, Document Cited by: Appendix D. CĆ©cillon et al. (2020) N. CĆ©cillon, V. Labatut, R. Dufour, and G. LinarĆØs WAC: a corpus of Wikipedia conversations for online abuse detection. In Proceedings of the Twelfth Language Resources and Evaluation Conference, N. Calzolari, F. BĆ©chet, P. Blache, K. Choukri, C. Cieri, T. Declerck, S. Goggi, H. Isahara, B. Maegaard, J. Mariani, H. Mazo, A. Moreno, J. Odijk, and S. Piperidis (Eds.), Marseille, France, p. 1382ā1390 (eng). External Links: Link, ISBN 979-10-95546-34-4 Cited by: Ā§2. Chen et al. (2022) Y. Chen, W. Fan, X. Xing, J. Pang, M. Huang, W. Han, Q. Tie, and X. Xu CPED: a large-scale chinese personalized and emotional dialogue dataset for conversational ai. arXiv preprint arXiv:2205.14727. External Links: Link Cited by: Ā§1, Ā§2. Desai et al. (2025) S. Desai, M. Dubiel, N. Zargham, T. Mildner, and L. Spillner Personas evolved: designing ethical LLM-based conversational agent personalities. In Proceedings of the 7th ACM Conference on Conversational User Interfaces (CUI ā25), External Links: Document Cited by: Ā§1, Evaluation Scope. Fawaz et al. (2026) K. Fawaz, R. Yi, O. Suciu, R. Khandelwal, H. Harkous, N. Taft, and M. Gruteser Text-based personas for simulating user privacy decisions. External Links: 2603.19791, Link Cited by: Ā§1, Evaluation Scope. Ferschke et al. (2012) O. Ferschke, I. Gurevych, and Y. Chebotar Behind the article: recognizing dialog acts in Wikipedia talk pages. In Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics, W. Daelemans (Ed.), Avignon, France, p. 777ā786. External Links: Link Cited by: Table 4, Appendix F, Appendix F, Appendix F, Ā§2, Ā§3.1, Ā§4, Informed Consent, footnote 11. Hua et al. (2018) Y. Hua, C. Danescu-Niculescu-Mizil, D. Taraborelli, N. Thain, J. Sorensen, and L. Dixon WikiConv: a corpus of the complete conversational history of a large online collaborative community. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, E. Riloff, D. Chiang, J. Hockenmaier, and J. Tsujii (Eds.), Brussels, Belgium, p. 2818ā2823. External Links: Link, Document Cited by: Ā§2, Ā§3.1, Informed Consent. Huang et al. (2023) Q. Huang, S. Fu, X. Liu, W. Wang, T. Ko, Y. Zhang, and L. Tang Learning retrieval augmentation for personalized dialogue generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, H. Bouamor, J. Pino, and K. Bali (Eds.), Singapore, p. 2523ā2540. External Links: Link, Document Cited by: Persona Conditioning. Jaidka et al. (2021) K. Jaidka, A. Ceolin, I. Singh, N. Chhaya, and L. Ungar WikiTalkEdit: a dataset for modeling editorsā behaviors on Wikipedia. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, K. Toutanova, A. Rumshisky, L. Zettlemoyer, D. Hakkani-Tur, I. Beltagy, S. Bethard, R. Cotterell, T. Chakraborty, and Y. Zhou (Eds.), Online, p. 2191ā2200. External Links: Link, Document Cited by: Ā§2. Jang et al. (2022) Y. Jang, J. Lim, Y. Hur, D. Oh, S. Son, Y. Lee, D. Shin, S. Kim, and H. Lim Call for customized conversation: customized conversation grounding persona and knowledge. In AAAI-22 Technical Tracks 10, Proceedings of the 36th AAAI Conference on Artificial Intelligence, AAAI 2022, p. 10803ā10812 (English). Note: Publisher Copyright: Copyright Ā© 2022, Association for the Advancement of Artificial Intelligence (w.aaai.org). All rights reserved.; 36th AAAI Conference on Artificial Intelligence, AAAI 2022 ; Conference date: 22-02-2022 Through 01-03-2022 External Links: Document Cited by: Ā§1, Ā§2. Johnson et al. (2024) I. Johnson, L. Kaffee, and M. Redi Wikimedia data for AI: a review of wikimedia datasets for NLP tasks and AI-assisted editing. In Proceedings of the First Workshop on Advancing Natural Language Processing for Wikipedia, L. Lucie-AimĆ©e, A. Fan, T. Gwadabe, I. Johnson, F. Petroni, and D. van Strien (Eds.), Miami, Florida, USA, p. 91ā101. External Links: Link, Document Cited by: Ā§2. Karan and Å najder (2019) V. M. Karan and J. Å najder Preemptive toxic language detection in Wikipedia comments using thread-level context. In Proceedings of the Third Workshop on Abusive Language Online, S. T. Roberts, J. Tetreault, V. Prabhakaran, and Z. Waseem (Eds.), Florence, Italy, p. 129ā134. External Links: Link, Document Cited by: Ā§2. Li et al. (2025) Z. Li, Y. Shi, Z. Liu, F. Yang, A. Payani, N. Liu, and M. Du Language ranker: a metric for quantifying llm performance across high and low-resource languages. In Proceedings of the Thirty-Ninth AAAI Conference on Artificial Intelligence (AAAI-25), p. 28186ā28194. Note: AAAIā25/IAAIā25/EAAIā25 Proceedings External Links: Document, Link Cited by: Ā§1. Lin (2004) C. Lin ROUGE: a package for automatic evaluation of summaries. In Text Summarization Branches Out, Barcelona, Spain, p. 74ā81. External Links: Link Cited by: Ā§4. Lin et al. (2021) Z. Lin, Z. Liu, G. I. Winata, S. Cahyawijaya, A. Madotto, Y. Bang, E. Ishii, and P. Fung XPersona: evaluating multilingual personalized chatbot. In Proceedings of the 3rd Workshop on Natural Language Processing for Conversational AI, A. Papangelis, P. Budzianowski, B. Liu, E. Nouri, A. Rastogi, and Y. Chen (Eds.), Online, p. 102ā112. External Links: Link, Document Cited by: Ā§2. Liu et al. (2025a) D. Liu, Z. Wu, D. Song, and H. Huang A persona-aware LLM-enhanced framework for multi-session personalized dialogue generation. In Findings of the Association for Computational Linguistics: ACL 2025, W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar (Eds.), Vienna, Austria, p. 103ā123. External Links: Link, Document, ISBN 979-8-89176-256-5 Cited by: Ā§1. Liu et al. (2025b) J. Liu, Y. Zhu, S. Wang, X. Wei, E. Min, Y. Lu, S. Wang, D. Yin, and Z. Dou LLMs + persona-plug = personalized LLMs. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar (Eds.), Vienna, Austria, p. 9373ā9385. External Links: Link, Document, ISBN 979-8-89176-251-0 Cited by: Persona Conditioning. Lotfi et al. (2023) E. Lotfi, M. De Bruyn, J. Buhmann, and W. Daelemans PersonalityChat: conversation distillation for personalized dialog modeling with facts and traits. In Proceedings of the Third Workshop on Natural Language Generation, Evaluation, and Metrics (GEM), S. Gehrmann, A. Wang, J. Sedoc, E. Clark, K. Dhole, K. R. Chandu, E. Santus, and H. Sedghamiz (Eds.), Singapore, p. 353ā371. External Links: Link Cited by: Ā§1, Ā§2. Malmqvist (2025) L. Malmqvist Sycophancy in large language models: causes and mitigations. In Intelligent Computing, K. Arai (Ed.), Cham, p. 61ā74. External Links: ISBN 978-3-031-92611-2 Cited by: Ā§2, Ā§5.2. Mansour et al. (2025) S. Mansour, L. Perelli, L. Mainetti, G. Davidson, and S. DāAmato PAARS: persona aligned agentic retail shoppers. In Proceedings of the 1st Workshop for Research on Agent Language Models (REALM 2025), E. Kamalloo, N. Gontier, X. H. Lu, N. Dziri, S. Murty, and A. Lacoste (Eds.), Vienna, Austria, p. 143ā159. External Links: Link, Document, ISBN 979-8-89176-264-0 Cited by: Ā§1, Ā§3.1, Evaluation Scope. MenĆ©ndez et al. (1997) M.L. MenĆ©ndez, J.A. Pardo, L. Pardo, and M.C. Pardo The jensen-shannon divergence. Journal of the Franklin Institute 334 (2), p. 307ā318. External Links: ISSN 0016-0032, Document, Link Cited by: Ā§4. Occhipinti et al. (2024) D. Occhipinti, S. S. TekiroÄlu, and M. Guerini PRODIGy: a PROfile-based DIalogue generation dataset. In Findings of the Association for Computational Linguistics: NAACL 2024, K. Duh, H. Gomez, and S. Bethard (Eds.), Mexico City, Mexico, p. 3500ā3514. External Links: Link, Document Cited by: Ā§1, Ā§2. Orlikowski et al. (2025) M. Orlikowski, J. Pei, P. Rƶttger, P. Cimiano, D. Jurgens, and D. Hovy Beyond demographics: fine-tuning large language models to predict individualsā subjective text perceptions. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar (Eds.), Vienna, Austria, p. 2092ā2111. External Links: Link, Document, ISBN 979-8-89176-251-0 Cited by: Ā§1, Ā§2. Panaretos and Zemel (2019) V. M. Panaretos and Y. Zemel Statistical aspects of wasserstein distances. Annual Review of Statistics and Its Application 6 (1), p. 405ā431. External Links: ISSN 2326-831X, Link, Document Cited by: Ā§4. Park et al. (2024) M. Park, S. Kim, S. Lee, S. Kwon, and K. Kim Empowering personalized learning through a conversation-based tutoring system with student modeling. In Extended Abstracts of the CHI Conference on Human Factors in Computing Systems, CHI EA ā24, New York, NY, USA. External Links: ISBN 9798400703317, Link, Document Cited by: Ā§1. Prabhakaran and Rambow (2016) V. Prabhakaran and O. Rambow A corpus of Wikipedia discussions: over the years, with topic, power and gender labels. In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LRECā16), N. Calzolari, K. Choukri, T. Declerck, S. Goggi, M. Grobelnik, B. Maegaard, J. Mariani, H. Mazo, A. Moreno, J. Odijk, and S. Piperidis (Eds.), Portorož, Slovenia, p. 2034ā2038. External Links: Link Cited by: Ā§2. Rupprecht et al. (2025) J. Rupprecht, L. Frƶhling, C. Wagner, and M. Strohmaier German general personas: a survey-derived persona prompt collection for population-aligned LLM studies. External Links: 2511.21722, Link Cited by: Ā§1, Evaluation Scope. Salemi et al. (2024) A. Salemi, S. Mysore, M. Bendersky, and H. Zamani LaMP: when large language models meet personalization. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), L. Ku, A. Martins, and V. Srikumar (Eds.), Bangkok, Thailand, p. 7370ā7392. External Links: Link, Document Cited by: Persona Conditioning. Santurkar et al. (2023) S. Santurkar, E. Durmus, F. Ladhak, C. Lee, P. Liang, and T. Hashimoto Whose opinions do language models reflect?. In Proceedings of the 40th International Conference on Machine Learning (ICML), External Links: Link Cited by: Ā§2. Searle (1969) J. R. Searle Speech acts: an essay in the philosophy of language. Cambridge University Press, Cambridge, UK. Cited by: Ā§1. Shanahan et al. (2023) M. Shanahan, K. McDonell, and L. Reynolds Role play with large language models. Nature 623, p. 493ā498. External Links: Document, Link Cited by: Ā§1. Smith et al. (2020) E. M. Smith, M. Williamson, K. Shuster, J. Weston, and Y. Boureau Can you put it all together: evaluating conversational agentsā ability to blend skills. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, D. Jurafsky, J. Chai, N. Schluter, and J. Tetreault (Eds.), Online, p. 2021ā2030. External Links: Link, Document Cited by: Ā§1, Ā§2. Takayama et al. (2025) J. Takayama, M. Ohagi, T. Mizumoto, and K. Yoshikawa Persona-consistent dialogue generation via pseudo preference tuning. In Proceedings of the 31st International Conference on Computational Linguistics, O. Rambow, L. Wanner, M. Apidianaki, H. Al-Khalifa, B. D. Eugenio, and S. Schockaert (Eds.), Abu Dhabi, UAE, p. 5507ā5514. External Links: Link Cited by: Persona Conditioning. Thakur et al. (2024) H. Thakur, E. Agrawal, and S. Mukund Personas within parameters: fine-tuning small language models with low-rank adapters to mimic user behaviors. In Adaptive Foundation Models: Evolving AI for Personalized and Efficient Learning, External Links: Link Cited by: Persona Conditioning. Tseng et al. (2024) Y. Tseng, Y. Huang, T. Hsiao, W. Chen, C. Huang, Y. Meng, and Y. Chen Two tales of persona in LLMs: a survey of role-playing and personalization. In Findings of the Association for Computational Linguistics: EMNLP 2024, Y. Al-Onaizan, M. Bansal, and Y. Chen (Eds.), Miami, Florida, USA, p. 16612ā16631. External Links: Link, Document Cited by: Ā§1. Wang and Cardie (2014) L. Wang and C. Cardie A piece of my mind: a sentiment analysis approach for online dispute detection. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), K. Toutanova and H. Wu (Eds.), Baltimore, Maryland, p. 693ā699. External Links: Link, Document Cited by: Ā§2. Wu et al. (2025) S. Wu, Y. Zhu, W. Hsu, M. Lee, and Y. Deng From personas to talks: revisiting the impact of personas on LLM-synthesized emotional support conversations. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, C. Christodoulopoulos, T. Chakraborty, C. Rose, and V. Peng (Eds.), Suzhou, China, p. 5439ā5453. External Links: Link, Document, ISBN 979-8-89176-332-6 Cited by: Ā§3.1. Wu et al. (2021) Y. Wu, X. Ma, and D. Yang Personalized response generation via generative split memory network. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, K. Toutanova, A. Rumshisky, L. Zettlemoyer, D. Hakkani-Tur, I. Beltagy, S. Bethard, R. Cotterell, T. Chakraborty, and Y. Zhou (Eds.), Online, p. 1956ā1970. External Links: Link, Document Cited by: Ā§1, Ā§2. Wulczyn et al. (2017) E. Wulczyn, N. Thain, and L. Dixon Ex machina: personal attacks seen at scale. In Proceedings of the 26th International Conference on World Wide Web, W ā17, Republic and Canton of Geneva, CHE, p. 1391ā1399. External Links: ISBN 9781450349130, Link, Document Cited by: Ā§2. Xu et al. (2022a) J. Xu, A. Szlam, and J. Weston Beyond goldfish memory: long-term open-domain conversation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), S. Muresan, P. Nakov, and A. Villavicencio (Eds.), Dublin, Ireland, p. 5180ā5197. External Links: Link, Document Cited by: Ā§1, Ā§2. Xu et al. (2022b) X. Xu, Z. Gou, W. Wu, Z. Niu, H. Wu, H. Wang, and S. Wang Long time no see! open-domain conversation with long-term persona memory. In Findings of the Association for Computational Linguistics: ACL 2022, S. Muresan, P. Nakov, and A. Villavicencio (Eds.), Dublin, Ireland, p. 2639ā2650. External Links: Link, Document Cited by: Ā§1, Ā§2. Xuan et al. (2025) W. Xuan, R. Yang, H. Qi, Q. Zeng, Y. Xiao, A. Feng, D. Liu, Y. Xing, J. Wang, F. Gao, J. Lu, Y. Jiang, H. Li, X. Li, K. Yu, R. Dong, S. Gu, Y. Li, X. Xie, F. Juefei-Xu, F. Khomh, O. Yoshie, Q. Chen, D. Teodoro, N. Liu, R. Goebel, L. Ma, E. Marrese-Taylor, S. Lu, Y. Iwasawa, Y. Matsuo, and I. Li MMLU-ProX: a multilingual benchmark for advanced large language model evaluation. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, C. Christodoulopoulos, T. Chakraborty, C. Rose, and V. Peng (Eds.), Suzhou, China, p. 1513ā1532. External Links: Link, Document, ISBN 979-8-89176-332-6 Cited by: Ā§1. Yamashita et al. (2023) S. Yamashita, K. Inoue, A. Guo, S. Mochizuki, T. Kawahara, and R. Higashinaka RealPersonaChat: a realistic persona chat corpus with interlocutorsā own personalities. In Proceedings of the 37th Pacific Asia Conference on Language, Information and Computation, C. Huang, Y. Harada, J. Kim, S. Chen, Y. Hsu, E. Chersoni, P. A, W. H. Zeng, B. Peng, Y. Li, and J. Li (Eds.), Hong Kong, China, p. 852ā861. External Links: Link Cited by: Ā§1, Ā§2. Zhang (2024) J. Zhang Guided profile generation improves personalization with large language models. In Findings of the Association for Computational Linguistics: EMNLP 2024, Y. Al-Onaizan, M. Bansal, and Y. Chen (Eds.), Miami, Florida, USA, p. 4005ā4016. External Links: Link, Document Cited by: Ā§3.1. Zhang et al. (2018a) J. Zhang, J. Chang, C. Danescu-Niculescu-Mizil, L. Dixon, Y. Hua, D. Taraborelli, and N. Thain Conversations gone awry: detecting early signs of conversational failure. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), I. Gurevych and Y. Miyao (Eds.), Melbourne, Australia, p. 1350ā1361. External Links: Link, Document Cited by: Ā§2. Zhang et al. (2018b) S. Zhang, E. Dinan, J. Urbanek, A. Szlam, D. Kiela, and J. Weston Personalizing dialogue agents: I have a dog, do you have pets too?. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), I. Gurevych and Y. Miyao (Eds.), Melbourne, Australia, p. 2204ā2213. External Links: Link, Document Cited by: Ā§2. Zhang et al. (2020) T. Zhang, V. Kishore, F. Wu, K. Q. Weinberger, and Y. Artzi BERTScore: evaluating text generation with BERT. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020, External Links: Link Cited by: Ā§4. Zhao et al. (2025) Z. Zhao, C. Vania, S. Kayal, N. Khan, S. B. Cohen, and E. Yilmaz PersonaLens: a benchmark for personalization evaluation in conversational AI assistants. In Findings of the Association for Computational Linguistics: ACL 2025, W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar (Eds.), Vienna, Austria, p. 18023ā18055. External Links: Link, Document, ISBN 979-8-89176-256-5 Cited by: Ā§1, Ā§2. Zheng et al. (2020) Y. Zheng, G. Chen, M. Huang, S. Liu, and X. Zhu Personalized dialogue generation with diversified traits. External Links: 1901.09672, Link Cited by: Ā§1, Ā§2. Appendix A Further Wiki-Talkie Details A.1 Statistics (a) de (b) en (c) es (d) fr (e) it Figure 6: Distributions of words per turn, turns per user, turns per thread, users per thread across the five languages. A.1.1 Distributions Figure 6 shows the distributions of the number of words per turn, turns per user, turns per thread, and users per thread across the five languages. These statistical properties exhibit remarkably consistent patterns across languages, suggesting that the conversational dynamics of the dataset are largely language-agnostic. Words per Turn The distribution is heavily right-skewed in every language, with the vast majority of turns containing fewer than 200 words, indicating that most conversational exchanges are brief while a small number of turns involve substantially longer responses. The 1,000-word cap applied uniformly across all languages, as threads containing longer turns were excluded due to their higher likelihood of noise. Turns per User This distribution is similarly right-skewed, with most users contributing only a handful of turns. The bulk of users fall below 50 turns across all languages, with the long tail suggesting a small cohort of highly active participants. The Italian and Spanish subsets exhibit a narrower spread, consistent with the smaller number of threads collected for these languages. Turns per Thread Threads are predominantly short, with the modal thread containing between 2 and 5 turns across all languages. This suggests that most conversations are transactional or goal-oriented rather than extended dialogues, with very few threads exceeding 10 turns. Users per Thread The vast majority of threads involve exactly 2 users, reflecting a typical dyadic conversation structure. Threads with 3 or more participants are rare across all languages, with the distribution dropping off sharply beyond 2 users. A.1.2 Users and Personas In Table 2, we show statistics on users and personas, with Figure 7 showing the corresponding word distributions for user pages and profiles. User pages are right-skewed across all languages, with most users having short pages and a minority producing substantially longer ones. Profiles follow a more bell-shaped distribution centered around 50ā100 words, reflecting the relatively standardized length of the persona descriptions we construct from user metadata, and confirming that our persona extraction distills user information into concise, consistent summaries regardless of the underlying page length. English users exhibit the richest descriptions, with both the longest average user pages and profile lengths (98.07 words), and near-complete coverage of contribution topics (100%), gender (67%), and profession (66%), reflecting the larger and more active English Wikipedia community. German users, while numerous, show lower rates of gender (36%) and profession (26%) disclosure despite high topic coverage (98.9%). Spanish, French, and Italian users display more modest attribute coverage, particularly for profession, reflecting sparser self-descriptions in these communities. The column reporting users with at least five turns indicates a core of highly active contributors in each language, ranging from approximately 30% (Spanish, Italian) to nearly 50% (German). Figure 8 shows the proportion of non-deducible or, less frequently, malformed behavioral trait values across all users, broken down by trait and language. Agreement Strategy, Consensus Orientation, and Disagreement Strategy consistently account for the largest shares across all languages, possibly due to the lower frequency of these behavioral aspects in user comments. Notably, Spanish exhibits the highest proportion in Agreement Strategy (ā¼ 0.50), while the remaining languages cluster more tightly around lower values. Traits such as Editing Behavior, Policy Usage, Tone Management, and Writing Style and Comment Structure show uniformly low proportions (ā¤ 0.10) across all languages. Overall, cross-language variation is modest, with no language systematically underperforming across all traits. (a) de (b) en (c) es (d) fr (e) it Figure 7: Word distributions for user pages and profiles. Total Users with Topics with Gender with Profession with >= 5 Turns Profile Length de 6,282 6,213 2,233 1,620 2,983 60.14 Ā± 27.42 en 4,580 4,580 3,062 3,038 1,874 98.07 Ā± 46.22 es 2,307 1,309 891 446 671 57.07 Ā± 30.73 fr 4,952 2,646 2,280 1,494 1,986 64.80 Ā± 35.23 it 1,850 1,089 882 381 678 56.06 Ā± 31.95 Table 2: Summary statistics of users, their attributes, and activity metrics (mean Ā± std for profile length). Figure 8: Proportion of unavailable trait values across all users, broken down by trait and language. A.2 Anonymization and Normalization All usernames in Wiki-Talkie are anonymized via deterministic hashing with a keyed hash function and salt, ensuring consistent mapping across the dataset while preventing reversal. Username occurrences are identified and replaced not only in structured metadata but also within comment text, where editors frequently mention one another through wikilink syntax or plain text references.1313 13 For risks associated with user re-identification, see Ethic Statement Section. Thread formatting is then cleaned and normalized: leading and trailing decorative characters (dashes, equals signs) are removed, HTML line breaks are converted to newlines, styling tags (<div>, <span>, <u>) are stripped while preserving content, <nowiki> and <code> tags are removed but enclosed text is retained, indentation markers are removed, and whitespace is collapsed and trimmed. MediaWiki templates and external links are preserved to maintain fidelity to the original discussions. The same normalization was applied also to articles. # Comments Avg. Prev. Len (Ā± Std) # Unique Users # Men # Women # Men Sent. # Women Sent. de 5,187 3.05 (Ā± 2.59) 564 195 55 1,638 1,212 en 8,658 3.39 (Ā± 3.05) 561 296 83 3,268 3,277 es 4,310 3.03 (Ā± 2.69) 298 83 20 1,129 458 fr 9,827 3.29 (Ā± 2.88) 682 250 61 2,535 2,566 it 4,706 3.19 (Ā± 2.92) 325 131 15 1,989 178 Table 3: Statistics of the sampled data used for experiments. Original Label Our Label CM, CW, CU, CS, CL, COBJ, CO Critique PSR Request PREF Reference PFC Commit PPC Report IP, IC Inform IS Ask ATT+ Positive ATT- Negative ATTP Partial Table 4: Mapping from original SEWD labels (Ferschke et al., 2012) to our dialog act taxonomy. # Comments Micro-F1 Ask Commit Critique Inform Negative Partial Positive Reference Report Request All 0.74 0.82 0.75 0.77 0.59 0.68 0.35 0.80 0.49 0.97 0.81 Three 0.74 0.84 0.79 0.78 0.58 0.65 0.19 0.79 0.55 0.98 0.83 Two 0.76 0.85 0.80 0.81 0.59 0.65 0.42 0.79 0.50 0.97 0.84 One 0.75 0.83 0.83 0.82 0.59 0.67 0.36 0.78 0.49 0.97 0.84 Table 5: Micro-F1 and F1 scores per label across different training settings. Appendix B Examples Tables 6ā10 present one representative example per language from the Wiki-Talkie dataset. Each table shows a discussion thread (including the article title, thread title, and the sequence of comments), along with the persona profiles of the participants. Due to space constraints, behavioral traits and articles are not shown. Language: German Page: Diskussion:Schwicheldt (Adelsgeschlecht) Thread: āBezugā User Profile 8bfc7f759c Der Benutzer ist seit dem 17. September 2004 bei Wikipedia registriert und feierte am 17. September 2025 seinen 21. Wiki-Geburtstag. Er hat insgesamt rund 284.000 Bearbeitungen vorgenommen, etwa 1100 Artikel verfasst und ungefƤhr 8400 Fotos hochgeladen. Er ist unter den 8000 aktivsten deutschsprachigen Wikipedianern. Der Benutzer erhielt 2016 den Denkmalpreis āSilberne Halbkugelā. Er ist Mitglied im Verein Wikimedia Deutschland ā Gesellschaft zur Fƶrderung Freien Wissens. Seine BeitrƤge konzentrieren sich auf Themen wie Bessarabien, den Wolfsburger und Helmstedter/Gifhorner Raum, Hannover, Hann. MĆ¼nden sowie ArchƤologie, Geschichte, Burgen, Schlƶsser, Kirchen und lokale DenkmƤler in Niedersachsen. 4f3806c190 Der Benutzer ist ein deutscher Muttersprachler mit fortgeschrittenen Englischkenntnissen. Er hat sich am 19. Dezember 2004 bei Wikipedia registriert und kĆ¼mmert sich hauptsƤchlich um historischen Bergbau, insbesondere um den Zwickauer Steinkohlenbergbau. Er hat zahlreiche Artikel zu bergbauhistorischen und bergbautechnischen Themen erstellt sowie zu verschiedenen anderen Themen wie Waffen, MilitƤrtechnik und Flugzeugen beigetragen. Author Comment 8bfc7f759c Besteht ein Bezug zum Ort [[Schwicheldt]]? 4f3806c190 Wenn man dem Ortsartikel glauben darf: ja. 8bfc7f759c Dann sollte das aber auch im Adelsartikel stehen. Table 6: German example: persona profiles and thread. Language: English Page: Talk:Cleveland Bay Thread: āMergeā User Profile 5f962a065d The user lives in Montana and owns horses. She has been on Wikipedia for a while and works as a small-town lawyer in private practice, which causes her availability to fluctuate. She occasionally writes about legal topics on Wikipedia but prefers other subjects since she does enough legal writing professionally. She also writes books and does freelance writing outside of Wikipedia. She is not compensated for her Wikipedia contributions. She takes photos of objects and locations in Montana, particularly for National Register of Historic Places listings, and uploads them to Commons. She can access documents at the Montana Historical Society. She has met other Wikipedians in person, helped facilitate editathons in Montana, and attended Wikipedia-related events such as WikiConference North America/USA. Her main contributions focus on horse-related articles and Montana-related topics including historic places. 95725d4ec5 The user is a woman whose main focus on Wikipedia is equine articles, although she also contributes to topics related to history, astronomy, books, endurance sports, the outdoors, and labor unions. She has been doing work on stub sorting, red link recovery, and other cleanup tasks. She is active on and off as a Good Article reviewer and serves as a delegate at Featured article review and as an administrator on Wikipedia. She is married. Author Comment 5f962a065d Iām only wondering if the Chapman horse contributed significantly to anything other than the Cleveland Bay, thus warranting its continued independent article. (Thinking of the Narragansett Pacer or Old English Black horse). Other than that, no problem with the merge, content seems similar in both. 95725d4ec5 Iāve never found a source that references the Chapman horse in relation to any breed other than the Cleveland Bay. All of the sources I combed through for this article seemed to think that the Chapman was basically the earliest incarnation of the Cleveland Bay, rather than a separate breed that they were descended from. So, I think that the merge is a safe bet, especially since there are no sources in the Chapman article that actually show why the author thought it was a separate breed. 5f962a065d Groovy. Table 7: English example: persona profiles and thread. Language: Spanish Page: DiscusiĆ³n:Calor especĆfico Thread: āNĆŗmero de citasā User Profile e1c14a406a El usuario expresa interĆ©s en tres Ć”reas principales: la lingĆ¼Ćstica y las lenguas del mundo, la fĆsica matemĆ”tica y la ingenierĆa, y la historia econĆ³mica, economĆa y sociologĆa. Tiene enlaces a sus pĆ”ginas de usuario en las Wikipedias en alemĆ”n, inglĆ©s, francĆ©s y portuguĆ©s. 94a136608d El usuario es un antiguo bibliotecario peruano que se identifica como miembro de ACAD. EstĆ” agradecido con la Wikipedia en espaƱol. Sus intereses incluyen Moquegua, las Olimpiadas y la QuĆmica. 4901fd71f4 El usuario es un hombre de Quito, Ecuador, que trabaja para el experimento CMS del CERN. Tiene un doctorado en FĆsica Experimental de PartĆculas. En el pasado dedicaba mucho tiempo a editar Wikipedia en espaƱol, pero actualmente lo hace cuando puede debido a la falta de tiempo libre. Ha contribuido a numerosos artĆculos relacionados con fĆsica, incluyendo mecĆ”nica cuĆ”ntica, teorĆa de la relatividad, fĆsica de partĆculas y electromagnetismo. Author Comment e1c14a406a A peticiĆ³n de otro wikipedista leĆ extensamente el artĆculo, fijĆ”ndome en especial en el nĆŗmero de afirmaciones que podrĆan requerir cita. Aunque las citas son algo deseable y casi siempre necesarias en este caso concreto no veo demasiadas afirmaciones extraordinarias o polĆ©micas que requieran citas. En un artĆculo sobre temas polĆ©micos las citas son mĆ”s necesarias, pero creer que todo artĆculo sea del tema que se requiere el mismo nĆŗmero de citas es errĆ³neo. En este caso el artĆculo trata de un tema de ciencia convencional muy asentado que podrĆamos considerar un tema ya muy asentado, en ese caso creo que al artĆculo no le faltan ya mĆ”s citas. Tener citas por tener citas, no garantiza la calidad o correcta referenciaciĆ³n del artĆculo, el nĆŗmero de citas debe depender el tipo de artĆculo, la temĆ”tica y lo polĆ©mico del tema. Mi apreciaciĆ³n es que en este caso el artĆculo tiene el nĆŗmero de citas adecuadas. En cualquier caso si alguien considera que no es asĆ deberĆa aƱadir QUĆ afirmaciones concretas considera que deben ser referenciadas. 94a136608d Hola, revisa arriba y estĆ”n las frases que necesitan referencias y tambiĆ©n lee [[WP:FP]]. 4901fd71f4 Hola Fideā¦ bueno he puesto referencias a las afirmaciones que dices que faltaban, me faltan solo las dos ultimas que no encuentro. Por el formato de las referencias, eso ya lo arregle. Espero que con eso sea suficiente. (Y vamos que no es un nominado para destacado!!) Saludos 94a136608d Disculpame si te parezco exigente, pero segĆŗn [[WP:SAB]]: cita|- Califica cuidadosamente. - Ā”Los estĆ”ndares de los ArtĆculos buenos son tan altos como aquellos del calificador mĆ”s indulgente!|[[WP:SAB]] Pero crĆ©eme, te va a salir un superBueno, de lo mejor, y si esas frases no fueran un poquito dudosas, ni las mencionarĆa. Bendiciones. 4901fd71f4 Si entiendo tu postura Fide, solo no entendi lo del superbueno??? Saludos y seguirĆ© buscando las referencias. 94a136608d Era solo un decir. Saludos Table 8: Spanish example: persona profiles and thread. Language: French Page: Discussion:Pfalz-ardenner/Bon article Thread: (untitled) User Profile 702383d34d Lāutilisateur a commencĆ© Ć contribuer sur WikipĆ©dia en dĆ©cembre 2006. Il se dĆ©crit comme un wikignome Ć ses heures perdues et travaille principalement sur des articles liĆ©s Ć la physique nuclĆ©aire, aux accĆ©lĆ©rateurs de particules, aux biographies de scientifiques et Ć lāIrlande. Il a significativement contribuĆ© Ć des articles sur la capture Ć©lectronique, la dĆ©sintĆ©gration du neutron libre et les noyaux p. d637e7f14a Lāutilisatrice est une femme inscrite sur WikipĆ©dia depuis 2007. Elle est une contributrice trĆØs active, ayant lancĆ© environ 230 propositions de labellisation dāarticles, dont 83% concernent le cheval. Elle a participĆ© Ć plusieurs publications scientifiques. Ses domaines de contribution principaux sont les crĆ©atures lĆ©gendaires (fĆ©es, lutins, malebestes), les chevaux (aspects culturels, sociologiques et gĆ©nĆ©tiques), et lāautisme. Elle sāintĆ©resse de plus en plus Ć la gĆ©nĆ©tique. Elle apprĆ©cie les tableaux prĆ©raphaĆ©lites. Depuis 2017, elle pratique la photographie animaliĆØre comme hobby, cherchant Ć photographier un maximum dāespĆØces et de races animales. Elle prĆ©fĆØre ĆŖtre tutoyĆ©e. 75d8423c61 Lāutilisateur est passionnĆ© de sport automobile, plus particuliĆØrement dāendurance. Il est nĆ© le jour du Grand Prix automobile du Portugal 1995. Il sāintĆ©resse aux histoires oubliĆ©es du sport automobile. Sur WikipĆ©dia, il contribue activement en crĆ©ant et amĆ©liorant de nombreux articles liĆ©s au sport automobile, notamment sur les automobiles de course, les courses dāendurance, les Ć©curies, les pilotes, ainsi que sur dāautres sujets comme la course au large, le football, lāhistoire militaire et lāarchitecture. Author Comment 702383d34d TrĆØs beau travail de recherche u-|user_d637e7f14a. Je nāai fait que survoler lāarticle mais une remarque me vient au passage. Ćtant donnĆ© que quasiment toutes les sources sont en anglais ou en allemand, dāoĆ¹ vient le nom en franƧais de cette race. Est-ce quāil est mentionnĆ© dans Guide des chevaux dāEurope qui est la seule source en franƧais ? Ou bien est-ce quāil sāagit de ta propre traduction ? Je demande cela car les seuls rĆ©sultats que je trouve sur le Web pour ce mot proviennent de WikipĆ©dia ou de clones. d637e7f14a Merci pour la question user_702383d34d. Le Guide des chevaux dāEurope a gardĆ© le nom allemand, ce qui est surprenant car toutes les autres races allemandes ont un nom traduit. Jāai tentĆ© une traduction littĆ©rale (dāabord Ā« du Palatinat Ā», puis Ā« palatin Ā») en dĆ©but de travail, mais je peux retourner Ć Pfalz-ardenner, qui effectivement est la seule rĆ©fĆ©rence attestĆ©e dans un bouquin en franƧais 702383d34d En fait, je ne sais pas comment on gĆØre Ƨa sur WikipĆ©dia habituellement. Ća ne me dĆ©range pas outre mesure de franciser le nom, mais Ć minima, je pense quāil faudrait indiquer dans une note que cāest une traduction Ā« personnelle Ā» (je nāai pas trop dāidĆ©e sur la meilleure faƧon de faire). Peut-ĆŖtre faut-il garder le nom en allemand et indiquer en note, ta traduction en franƧais. Peut-ĆŖtre que les autres votants ont un avis sur la question. 75d8423c61 Si le titre de lāarticle est une traduction personnelle, alors il vaut mieux renommer lāarticle. Pour garder le [[WP:PMS]]. Peut-ĆŖtre Ć©crire la traduction dans une note oĆ¹ la traduire dans lāintro. Mais il faut renommer 75d8423c61 Merci, cāest mieux comme Ƨa ! Table 9: French example: persona profiles and thread. Language: Italian Page: Discussione:Nevicata del 1985 Thread: āLeggiamo lāarticoloā User Profile 6a9673768 Lāutente si ĆØ registrato su Wikipedia lā8 agosto 2006. Contribuisce principalmente nellāambito delle voci scientifiche, con particolare attenzione alle Scienze della Terra, e interviene anche su voci riguardanti storiografia, filosofia e cinema, che sono sue vecchie passioni. Ć anche appassionato di fotografia e inserisce immagini originali di sua produzione. Partecipa al Progetto sulle persecuzioni, deportazioni e crimini del periodo nazi-fascista, in particolare sul tema del Negazionismo dellāOlocausto. Ć amministratore di Wikipedia dal 2007. 677e695641 Lāutente ĆØ un uomo over 60, curioso e intellettualmente aperto a nuovi apprendimenti. Si appassiona in particolare a temi concernenti diritto, letteratura, musica e scienze politiche. Si ĆØ registrato su Wikipedia verso la fine di febbraio 2006. Ć laureato in Giurisprudenza. Vive a La Spezia. Su Wikipedia ha creato o modificato consistentemente voci riguardanti societĆ , diritto, economia, storia, istituzioni, politica, musica, letteratura, arte, religione, filosofia e sport. 1363852f3c Lāutente ĆØ attivo su Wikipedia dal 21 agosto 2006, quando ha effettuato la sua prima modifica. Lā11 febbraio 2008 ĆØ stato eletto amministratore. Ha raggiunto diversi traguardi di contributi: 100 edit il 24 marzo 2007, 500 il 19 giugno 2007, 1000 il 21 luglio 2007, 2000 lā8 agosto 2007, 4000 il 22 settembre 2007, 5000 il 17 ottobre 2007, 7500 il 18 gennaio 2008, 10000 il 13 aprile 2008 e 20000 il 16 aprile 2009. Ha creato 100 nuove voci entro il 21 ottobre 2007. Contribuisce principalmente ad articoli sulla musica punk, hardcore punk, gruppi musicali e album, con particolare attenzione a band come Less Than Jake, Dropkick Murphys, Misfits, Sick of It All, Weezer e alla scena hardcore svedese. Author Comment 6a9673768 Nellā ultimo intervento user_677e695641 ha richiesto di inserire una citazione giaā presente da almeno una settimana nelle fonti/bibliografia. Questo mi fa sospettare che i suoi interventi e commenti non sia preceduti da una corretta/completa lettura della voce. Lascio a chi legge i commenti del caso. 677e695641 Hai ragione, scusa: avevo letto piĆ¹ volte lāarticolo ma non avevo ancora aperto il link a quellāarticolo (Nimbus) che usa questo appellativo di ānevicata del secoloā (mi pare lāunico caso, ma comunque esiste), anche se di norma con nevicata del secolo lāopinione pubblica pensa a quella del ā56. 1363852f3c Visto che lāarticolo linkato parla di āNevicata del Secoloā riferendosi perĆ² solo alla cittĆ di Milano o, comunque, solo al settentrione, potremmo, nel caso in cui si decidesse per lāunificazione delle pagine, mantenere la nomencaltura āNevicata del Secoloā per lāeventuale paragrafo di approfondimento sulla situazione nel Nord Italia. qualcosa del tipo La āNevicata del Secoloā di Milano. Che ne dite? 677e695641 Il punto ĆØ che mi convince poco lāuso di iperboli e di espressioni altisonanti, tipo ānevicata del secoloā. Inoltre, pare che di nevicate del secolo ce ne siano state piĆ¹ di una (leggi la disambigua [[Nevicata del secolo]]). 6a9673768 La disambigua di cui sopra eā stata creata da user_677e695641 appositamente per poter affermare che ci sono piuā nevicate secolari, il che eā una ovvietaā e veritaā statistica: ogni zona dā Italia ha avuto la sua nevicata secolare. Riguardo alla iperbole si rimanda allā articolo citato quiā sopra 677e695641 La disambigua non ĆØ stata creata appositamente per poter affermare che ci sono piuā nevicate secolari, ma proprio perchĆØ lāesistenza di piĆ¹ fenomeni chiamati ānevicata del secoloā - come lo stesso user_6a9673768 riconosce - ĆØ una ovvietĆ e veritĆ statistica: caso classico in cui una disambigua si impone. Suggerirei, se mi ĆØ consentito, di usare toni meno polemici e piĆ¹ legati al merito delle questioni. Table 10: Italian example: persona profiles and thread. Appendix C Prompts We report here the prompts used in this study. Figure 9 shows the prompt for thread revision (Section 3.1). Figures 10, 11, and 13 present the prompts for persona construction (Section 3.1). Figure 12 shows the prompt used for dialog act evaluation (Section 4). Figures 14, 15, 16, and 17 present the prompts used for the conditioning methods (Section 4). All prompts follow a two-part structure: a system prompt establishing the modelās role and output constraints, and a user prompt providing task instructions, definitions, few-shot exemplars where applicable, and the input to process. Placeholders in braces are filled at inference time. System Prompt You are an expert in analyzing Wikipedia Talk pages written in language. You deeply understand: ā¢ how Wikipedia discussions are structured ā¢ how user signatures and timestamps work ā¢ how informal, unstructured talk-page conversations appear You must correct errors produced by automatic splitting of Wikipedia discussions. You always follow the output format rules exactly and never add explanations. User Prompt I am processing Wikipedia Talk page conversations written in language. These conversations were automatically split into comments using heuristics (mainly assigning each comment to the last detected user signature). Because Talk pages are informal and unstructured, these heuristics sometimes produce incorrect splits. You are given JSON data that includes: ā¢ thread: the automatically split discussion (may contain errors) ā¢ free_form_thread: the original unstructured conversation (ground truth) Your task is to determine whether thread correctly reflects the structure and speaker attribution implied by free_form_thread, and to fix it if possible. ### DECISION PROCEDURE 1. Compare thread against free_form_thread. 2. If speaker attribution or comment boundaries cannot be determined with high confidence, output exactly: 3. Otherwise, apply only the minimal necessary fixes to the thread field. 4. If no fixes are required, return the input thread field unchanged. ### OUTPUT RULES 1. Output ONLY valid JSON. 2. Output must contain ONLY the thread field. If attribution is too ambiguous, output exactly . 3. Modify the thread field ONLY IF a fix is clearly justified. 4. Do NOT reformat text, normalize whitespace, alter punctuation, or modify timestamps. 5. Preserve all comment text verbatim except when removing a clearly identifiable trailing user signature. 6. Do NOT add, infer, guess, merge, split, reorder, or rewrite content except as explicitly permitted below. ### TYPES OF ERRORS TO FIX 1. Merged comments (missing signatures) Multiple distinct usersā comments may be merged into a single entry. Fix ONLY if distinct speakers are clearly identifiable and comment boundaries are unambiguous. If signatures are missing or attribution is uncertain, output . [ Few-shot examples: one merge fix, two ambiguous cases outputting ] 2. Post-signature edits (post scriptum) Text added after a user signature should belong to the preceding comment, not the following one. [ Few-shot example ] 3. Trailing signature not removed If a comment still contains its own signature at the end, remove only the signature text. [ Few-shot example ] 4. Already correct thread If thread already matches the structure implied by free_form_thread, return it unchanged. [ Few-shot example ] ### ACTUAL INPUT INPUT: input_row OUTPUT: Figure 9: Prompt template for thread split revision. Few-shot example content is summarized for brevity. System Prompt You extract explicit personal information from Wikipedia user pages written in language. Follow the user prompt exactly. If information is missing or unclear, use āunknownā. Always output exactly one valid JSON object and nothing else. Do not add explanations, comments, or formatting outside the JSON. User Prompt You are given the full text of a Wikipedia user page written in language. The page may contain explicit self-declared information about the user (background, personal traits, interests), or it may contain only irrelevant content or no personal information at all. ## TASK 1. Determine whether the page contains any explicit personal information about the user. 2. If yes: a. Write a natural-language profile containing that information. b. Extract the information into the structured fields below. ## DEFINITION Personal information includes only what the user explicitly states about themselves, such as: ā¢ Background or biographical details ā¢ Gender or age ā¢ Location (origin or residence, if explicitly stated) ā¢ Languages spoken ā¢ Education and profession ā¢ Interests or hobbies ā¢ Wikipedia-related activity (e.g., registration year, preferred topics, articles edited or created) ## RULES (STRICT) ā¢ Use only information explicitly stated on the user page; do NOT invent anything. ā¢ Do NOT include names, usernames, email addresses, or contact details in the profile. ā¢ Information in structured fields must be in English, except profile, which must be written in language. ā¢ If a field is missing or unclear, use āunknownā. ā¢ If no personal information is present, set relevant_page to false. ā¢ Output exactly one valid JSON object and nothing else. ## FIELDS TO OUTPUT relevant_page: Boolean; true if any explicit personal information is present. profile: Natural-language description in language, third person, no names, no full birth dates (birth years only). "" if relevant_page is false. gender: "man", "woman", "non-binary", "unknown". Infer from explicit statements or grammatical gender cues; do NOT infer from names. age_range: "under 18", "18--30", "31--40", "41--60", "over 60", "unknown". Only if explicitly stated. location: Country or city explicitly stated as place of residence; "unknown" if absent. known_languages: Languages explicitly stated; names in English. [] if absent. education_level: "no formal education", "high school", "undergraduate", "graduate", "postgraduate", "unknown". Only if explicitly stated. study_field: Field associated with education, in English; "unknown" otherwise. profession: Explicitly stated profession, in English; "unknown" if absent. hobbies: Explicitly stated recreational activities, in English. [] if absent. stated_interests: Topics explicitly declared as interests or stated editing focus, in English. [] if absent. registration_year: Wikipedia registration year if explicitly stated; "unknown" otherwise. contribution_topics: Abstract topic categories derived from article lists explicitly present on the page, in English. [] if absent. ## EXAMPLE INPUT: user_page_true OUTPUT: profile_true INPUT: user_page_false OUTPUT: profile_false ## ACTUAL INPUT INPUT: input_row OUTPUT: Figure 10: Prompt template for user page profiling. System Prompt You are an AI assistant specialized in analyzing Wikipedia editor behavior. Your task is to extract editorial decision policies and interactional style traits from a userās comments on Wikipedia Talk pages written in language. Focus on observable, repeatable patterns in: ā¢ how the user argues and frames disagreement ā¢ how the user moderates or enforces norms ā¢ how the user structures Talk Page comments ā¢ how the user applies Wikipedia policies and guidelines ā¢ interactional style traits (e.g., hedging vs. assertiveness, politeness markers, escalation threshold) ā¢ writing style features that influence discussion dynamics (e.g., formal vs. informal register, use of emphasis such as bold or italics, structured formatting, rhetorical questions, or inclusion of emojis or informal symbols) Base all conclusions strictly on explicit evidence in the provided data. Do not generalize beyond recurring, observable patterns. If evidence is missing, conflicting, or insufficient, set "value" to ["Not-deducible"]. Always output exactly one valid JSON object and nothing else. Do not include explanations or formatting outside the JSON. User Prompt (1/2) You are given a set of Wikipedia Talk threads containing comments in language authored by the user username. <Comments> threads </Comments> ## TASK Analyze the data to extract usernameās editorial decision policies and interactional style traits, capturing how they typically behave and contribute in Wikipedia discussions. Focus on how username: ā¢ frames disagreements ā¢ applies policies ā¢ structures arguments ā¢ interacts with other editors ā¢ seeks (or resists) consensus ## STRICT RULES ā¢ Base your analysis only on observable evidence from the provided comments authored by username. Comments from other users are context only. ā¢ Do NOT invent motivations, beliefs, or background. ā¢ Do not generalize beyond recurring, observable patterns. If evidence is missing, conflicting, or insufficient, set "value" to ["Not-deducible"]. ā¢ Output must be exactly one valid JSON object. ā¢ Do NOT add any text outside the JSON. ## EDITORIAL BEHAVIORAL FEATURES TO EXTRACT For each field, provide: ā¢ "value": a list of values selected from the allowed options. You can select one or multiple options. ā¢ "reasoning": brief, evidence-based justification (2ā3 sentences max). ### 1. Tone Management How the user manages tone in discussions. Allowed values: Neutral and procedural, Conciliatory and polite, Firm and directive, Confrontational, Not-deducible. ### 2. Consensus Orientation How the user approaches consensus-building. Allowed values: Actively seeks compromise, Willing to compromise with conditions, Defends position until consensus shifts, Prioritizes enforcement over consensus, Not-deducible. ### 3. Agreement Strategy How the user typically responds when agreeing. Allowed values: Explicit endorsement with reasoning, Brief affirmation, Conditional agreement, Agreement with modification proposal, Neutral acknowledgment, Not-deducible. ### 4. Disagreement Strategy How the user typically responds when disagreeing and critiquing work done by other users or their opinions. Allowed values: Clarification-seeking, Policy-based objection, Assertion of violation, Compromise-oriented, De-escalation-focused, Not-deducible. Figure 11: Prompt template for editorial behavior extraction (system prompt and user prompt, part 1). System Prompt You are an expert annotator who tags discussion turns in Wikipedia talk pages in language with dialog act labels. Each turn can have multiple labels. You must read the conversation context carefully and assign all relevant labels to a given target turn. Do not explain your reasoning, return only the labels as a comma-separated list. Be precise and consistent with the label definitions. User Prompt You are provided with a comment in language to label with dialog acts. It can receive multiple labels. In the following, you can find the label set and their definitions. Return the label names EXACTLY as written. ## Labels CRITIQUE Use this when a comment explicitly criticizes something. This can include the current article content, prior edits, structure, style, accuracy, completeness, neutrality, or clarity. Key signals: points out errors, says content is lacking, inaccurate, poorly structured, non-neutral. Example: āThis section on the filmography is incomplete and lacks reliable sources.ā REQUEST Use this when a comment contains a suggestion, recommendation, or explicit request for action. Key signals: suggests, recommends, asks to add/remove/change something. Example: āWe should add a citation for this paragraph.ā REFERENCE Use this when a comment points to an external source or Wikipedia page without requesting action. Key signals: mentions [[Wikipedia page]], cites URL, provides a pointer to a guideline. Example: āSee [[WP:NPOV]] for more information.ā COMMIT Use this when an editor states an intention to act in the future. Key signals: āI willā¦ā, āI plan toā¦ā, āI intend toā¦ā. Example: āI will update the article tomorrow.ā REPORT Use this when an editor reports that an action has already been performed. Key signals: āI addedā¦ā, āI changedā¦ā, āI correctedā¦ā, done. Example: āI just added the missing reference.ā INFORM Use this label when a comment provides information, factual details, or corrections to other editorsā understanding. Key signals: shares knowledge, clarifies, corrects misinformation. Example: āThe population of Paris is 2.1 million according to the latest census.ā ASK Use this when a comment asks for information or clarification. Key signals: āDoes anyone knowā¦?ā, āCan someone explainā¦?ā, āWhat is the sourceā¦?ā Example: āCan someone provide a source for this statement?ā POSITIVE Use this when a comment expresses agreement, thanks, congratulations, or appreciation toward another editor. Key signals: agrees, thanks, congratulates, supports another editorās point. Example: āThanks for adding this, I agree with your suggestion.ā NEGATIVE Use this when a comment expresses disagreement, rejection, or complaint toward another editor or their edit proposals. Key signals: disagrees, rejects (notdone), complains, criticizes another editor (not the article). Example: āI disagree with this change; it introduces errors.ā PARTIAL Use this when a comment both acknowledges and partially rejects another editorās point (a mixture of POSITIVE and NEGATIVE). Key signals: agrees with some points, but disagrees or partially rejects others. Example: āI agree with your point, but I think this part still needs work.ā ## Task ā¢ Assign all relevant labels to ācomment_to_labelā ā¢ You are provided with up to the two most recent previous comments and the thread title for context ā¢ Return them as a comma-separated list ā¢ Do not include explanations, reasoning, or additional text exemplars ### Data data_to_label ### Labels Figure 12: Prompt template for dialog act labeling. User Prompt (2/2) ### 5. Informing Strategy What the user provides in support of their claims (knowledge, sources, or article excerpts). Allowed values: Personal knowledge or expertise claims, Citation of Wikipedia policies or guidelines, External reliable sources, Article text excerpts or diffs, Logical or procedural reasoning, Not-deducible. ### 6. Policy Usage How the user uses Wikipedia policies and guidelines. Allowed values: Frequent explicit citation, Occasional citation, Implicit norm-based reasoning, Rare or no policy reference, Not-deducible. ### 7. Editing Behavior How the user reports edits or willingness to edit or request edits. Allowed values (select all that apply): Proposes specific edits, Performs edits and reports them, Requests others to edit, Recommends reverts or closures, Avoids editing and focuses on discussion, Not-deducible. ### 8. Writing Style and Comment Structure Typical structure of the userās Talk Page comments and style. Allowed values (select all that apply): Long, detailed comments, Short, concise comments, Bullet-pointed structure, Question-led structure, Action-oriented closings, Informal conversational tone, Formal procedural tone, Includes emojis or informal symbols, Polite or hedging language, Assertive or directive language, Not-deducible. ## OUTPUT FORMAT Provide ONLY valid JSON using the following structure: output_template Figure 13: Prompt template for editorial behavior extraction (user prompt, part 2). System Prompt You are participating as user, a Wikipedia editor, in a Wikipedia Talk Page discussion. You are proficient with Wikipedia guidelines and formatting conventions, including: ā¢ External links: [https://example.com Description] ā¢ Internal links: [[Article title]] or [[Article title|display text]] ā¢ Policy and guideline references: [[WP:NPOV]], [[WP:V]], [[WP:OR]], [[WP:RS]] ā¢ Templates: template name ā¢ Emphasis markup (āboldā, āitalicsā) ā¢ List formatting: * for bulleted lists, # for numbered lists You generate the next comment in language for the discussion, using Wikipedia formatting where appropriate. User Prompt You are provided with the article āarticle_titleā and a discussion titled āthread_titleā about it. ## Wikipedia Article article ā ## Target Discussion previous_comments Given the target discussion above, generate the next comment in language at timestamp timestamp as user, coherently and logically continuing the discussion so far. Do not refer to being an AI assistant, do not include meta-commentary, and do not add signatures (ā¼ ā¼ ā¼ ā¼ ) or timestamps. Output only the comment itself. Figure 14: Prompt template for base. System Prompt You are participating as user, a Wikipedia editor, in a Wikipedia Talk Page discussion. You are proficient with Wikipedia guidelines and formatting conventions, including: ā¢ External links: [https://example.com Description] ā¢ Internal links: [[Article title]] or [[Article title|display text]] ā¢ Policy and guideline references: [[WP:NPOV]], [[WP:V]], [[WP:OR]], [[WP:RS]] ā¢ Templates: template name ā¢ Emphasis markup (āboldā, āitalicsā) ā¢ List formatting: * for bulleted lists, # for numbered lists You generate the next comment in language for the discussion, using Wikipedia formatting where appropriate. You must adapt your expertise, reasoning, communication style, and tone to fully align with your profile below. This includes tailoring both the level and domain of knowledge demonstrated in your responses to match the documented attributes and characteristics in your profile. ## Profile user_information User Prompt You are provided with the article āarticle_titleā and a discussion titled āthread_titleā about it. ## Wikipedia Article article ā ## Target Discussion previous_comments Given the target discussion above, generate the next comment in language at timestamp timestamp as user, coherently and logically continuing the discussion so far. The comment must reflect the documented attributes in your profile. Do not refer to being an AI assistant, do not include meta-commentary, and do not add signatures (ā¼ ā¼ ā¼ ā¼ ) or timestamps. Output only the comment itself. Figure 15: Prompt template for profile. System Prompt You are participating as user, a Wikipedia editor, in a Wikipedia Talk Page discussion. You are proficient with Wikipedia guidelines and formatting conventions, including: ā¢ External links: [https://example.com Description] ā¢ Internal links: [[Article title]] or [[Article title|display text]] ā¢ Policy and guideline references: [[WP:NPOV]], [[WP:V]], [[WP:OR]], [[WP:RS]] ā¢ Templates: template name ā¢ Emphasis markup (āboldā, āitalicsā) ā¢ List formatting: * for bulleted lists, # for numbered lists You generate the next comment in language for the discussion, using Wikipedia formatting where appropriate. You must adapt your reasoning, communication style, and tone to fully align with the interaction traits provided below. This includes tailoring your interactions and communication to match the documented traits. ## Traits user_information User Prompt You are provided with the article āarticle_titleā and a discussion titled āthread_titleā about it. ## Wikipedia Article article ā ## Target Discussion previous_comments Given the target discussion above, generate the next comment in language at timestamp timestamp as user, coherently and logically continuing the discussion so far. The comment must reflect your documented interaction traits, both in editorial decisions and in interaction style. Do not refer to being an AI assistant, do not include meta-commentary, and do not add signatures (ā¼ ā¼ ā¼ ā¼ ) or timestamps. Output only the comment itself. Figure 16: Prompt template for behavior (user_information without reasoning) and behavior+r (user_information with reasoning). System Prompt You are participating as user, a Wikipedia editor, in a Wikipedia Talk Page discussion. You are proficient with Wikipedia guidelines and formatting conventions, including: ā¢ External links: [https://example.com Description] ā¢ Internal links: [[Article title]] or [[Article title|display text]] ā¢ Policy and guideline references: [[WP:NPOV]], [[WP:V]], [[WP:OR]], [[WP:RS]] ā¢ Templates: template name ā¢ Emphasis markup (āboldā, āitalicsā) ā¢ List formatting: * for bulleted lists, # for numbered lists You generate the next comment in language for the discussion, using Wikipedia formatting where appropriate. You must adapt your expertise, reasoning, communication style, and tone to be consistent with your own comments across other threads. User Prompt You are provided with the article āarticle_titleā and a discussion titled āthread_titleā about it. ## Wikipedia Article article ā In the following section you can find your history, that is, other threads where you wrote comments as user. ## User History history ā ## Target Discussion previous_comments Given the target discussion above, generate the next comment in language at timestamp timestamp as user, coherently and logically continuing the discussion so far. The comment must reflect the expertise, reasoning, communication style, and tone demonstrated in your user history. Do not refer to being an AI assistant, do not include meta-commentary, and do not add signatures (ā¼ ā¼ ā¼ ā¼ ) or timestamps. Output only the comment itself. Figure 17: Prompt template for history. Appendix D Data Sample Table 3 reports statistics of the sampled data from Wiki-Talkie. English and French are the largest splits, with roughly 9ā10k comments each. German, despite starting from the same number of initial threads (5,000), yielded fewer turns after subsampling. Spanish and Italian are roughly half the size of English and French. Thread length is consistent across languages, averaging around 3 previous comments per turn. All languages exhibit a strong male skew in user distribution, consistent with the well-documented gender gap in Wikipedia editorship Cabrera et al. (2018). The skew is most pronounced in Italian and Spanish, where the limited pool of female editors made gender balancing impossible. English and French are the most balanced: in the French split, for instance, male and female editors contribute nearly equal numbers of sentences despite a skewed ratio of unique users. Appendix E Generation Details All next-turn generations were produced using a cloud-based batch inference service. Generations were conducted with a temperature of 00 to ensure deterministic outputs, and a maximum of 1,0241,024 tokens were generated per response. To accommodate the context window constraints of each model, input prompts were dynamically truncated when necessary: if the combined length of the system and user prompts exceeded the available input budget (i.e., the modelās context window minus the reserved generation tokens), the article body was progressively shortened by iteratively reducing it to 90% of its current length until the prompt fit within the budget, with a <article_truncated> marker appended to signal the truncation. For the history method, if the full history caused the prompt to exceed the context window, the number of historical threads was progressively reduced until the prompt fit; instances where the resulting history fell below the minimum threshold of 55 threads were flagged as underrepresented. Prompts that could not be reduced to fit within the context window even after exhausting all truncation strategies were discarded. Appendix F Validation of Acts Annotation To validate our LLM-based annotation approach, we rely on the SEWD dataset Ferschke et al. (2012). SEWD provides act annotations for English Wikipedia talk page discussions, a setting analogous to Wiki-Talkie corpus. In SEWD, the reconciled labels of two trained annotators serve as the gold standard. We use these reconciled labels to evaluate the quality of our LLM-as-annotator pipeline (the extractor model). For clarity, we rename and merge some labels to reduce granularity; the full mapping is reported in Table 4. However, the format of SEWD differs substantially from Wiki-Talkie. In SEWD, label spans are assigned directly to raw comments within a thread, without prior splitting or cleaning, whereas in our scenario we aim to provide the LLM with preceding comments as conversational context and a single target reply to be labelled. To bridge this gap, we first convert SEWD into the Wiki-Talkie format, structuring each instance as a target comment to be annotated along with its preceding comments as context, and aligning the corresponding SEWD labels. This conversion inevitably introduces noise; we therefore manually clean the resulting instances and discard those that could not be reliably mapped. From the original 1,500 threads in SEWD, this procedure yields 735 usable threads, which constitute our validation set. Since replies in talk page discussions typically do not refer to much earlier context, we experiment with varying amounts of preceding conversational history: all previous comments, up to three, up to two, or only the immediately preceding comment. The prompt provided to the LLM includes explanations of each act label together with illustrative English examples drawn from SEWD. Table 5 reports the Micro-F1 and per-label F1 scores across these settings. Performance is stable across context windows, with the two-comment setting achieving the highest Micro-F1 (0.77). We therefore adopt this configuration for all subsequent annotation. Notably, these scores are very close to the human inter-annotator agreement reported in Ferschke et al., 2012 (Micro-F1 = 0.79), even though the two evaluation sets are not perfectly overlapping: ours is the product of format conversion and manual filtering, whereas the figure in Ferschke et al., 2012 results from cross-validation over the entire SEWD corpus. Nevertheless, this proximity strongly suggests that the LLM approximates human-level labeling quality. Examining the per-label scores, the most challenging categories (F1 << 0.70) are in order Partial, and Reference, Negative, Inform. To better understand these discrepancies, one of the authors of this paper manually inspected 20 instances per class where the LLM prediction and the ground truth disagreed. In more than half of the cases, the author judged the LLM label to be more appropriate than the gold label. This pattern may stem from a mismatch between how the act categories are explained in our prompt and how they were originally operationalized in Ferschke et al., 2012 (which is itself likely connected to the relatively low inter-annotator agreement reported in that work, arguably a consequence of insufficiently precise annotation guidelines). Figure 18: Proportion difference between human and agent dialog acts split across genders. Positive values indicate human over-production; negative values indicate agent over-production. Appendix G Full Results Tables 11, 12 and Tables 13, 14 report full results for text similarity and interaction behavior evaluation, respectively. Average BERTScore and ROUGE-L scores are discussed in Section 5.1, while average dialog act and sentiment scores are discussed in Section 5.2. Figure 18 reports the proportion difference between human and agent dialog acts split by gender, and Figure 19 shows the distributions of sentiment scores for human and agent comments, discussed in Section 5.2. In the following, we detail text length and act density ratio, which are not discussed in Section 5. Text Length Across all conditions, agent-generated comments are substantially longer than human comments. Human comments average approximately 48ā78 words depending on the language (de being the shortest at ā¼48 48 words and es the longest at ā¼78 78), while agent output consistently exceeds these figures, often by a wide margin. Notably, the history method yields the shortest agent comments across all modelsāfor instance, Model A drops to ā¼88 88 words on average, its closest approximation to human lengthāsuggesting that exposure to real interaction excerpts helps calibrate verbosity. Conversely, the behavior and profile methods tend to inflate output length, particularly for Model D and Model B. Interestingly, es also elicits slightly longer agent comments than other languages across all conditioning methods, mirroring the fact that human comments in es are themselves the longest. Act Density Ratio The act density ratio is consistently below 1.0 across all conditions, indicating that agent-generated comments are uniformly richer in communicative intents than their human counterparts. The lowest ratios are observed under the behavior method, where agents appear to enact the prescribed behavioral traits, which are dense in prescriptive interactional information. By contrast, when grounded with real examples via history, the ratio rises above 0.7 on average across all models and languages, with the exception of Model B, suggesting that concrete exemplars help bring act density closer to human levels. (a) Model A under base method. (b) Model B under base method. (c) Model B under history method. Figure 19: Sentiment scores distributions for human agent comments on en for selected models and conditioning methods. Setup Lang Model Avg Human Words Avg Agent Words BERT ā rBERT ā ROUGE ā base de Model A 47.9 (56.3) 80.2 (34.2) 0.849 0.065 0.096 Model B 47.7 (56.3) 205.1 (74.3) 0.822 -0.102 0.061 Model C 47.9 (56.3) 87.7 (32.5) 0.847 0.051 0.088 Model D 47.9 (56.3) 121.7 (47.4) 0.837 -0.013 0.077 Avg models 47.8 Ā± 0.1 123.7 Ā± 57.2 0.839 Ā± 0.012 0.000 Ā± 0.076 0.081 Ā± 0.015 en Model A 62.4 (78.7) 105.4 (54.3) 0.854 0.013 0.124 Model B 62.4 (78.7) 251.1 (109.2) 0.832 -0.140 0.085 Model C 62.4 (78.7) 105.7 (46.7) 0.852 -0.001 0.120 Model D 62.4 (78.7) 132.2 (54.7) 0.843 -0.065 0.101 Avg models 62.4 Ā± 0.0 148.6 Ā± 69.5 0.845 Ā± 0.010 -0.048 Ā± 0.070 0.107 Ā± 0.018 es Model A 77.5 (96.5) 113.8 (49.6) 0.852 0.050 0.132 Model B 77.5 (96.8) 270.0 (92.8) 0.829 -0.099 0.094 Model C 77.5 (96.5) 116.4 (52.1) 0.851 0.039 0.129 Model D 77.5 (96.5) 152.4 (60.2) 0.842 -0.014 0.115 Avg models 77.5 Ā± 0.0 163.1 Ā± 73.4 0.844 Ā± 0.011 -0.006 Ā± 0.068 0.118 Ā± 0.017 fr Model A 63.9 (77.2) 100.7 (42.5) 0.851 0.061 0.117 Model B 63.9 (77.2) 238.0 (82.1) 0.828 -0.086 0.083 Model C 63.9 (77.2) 107.6 (45.6) 0.849 0.048 0.110 Model D 63.9 (77.2) 141.6 (55.9) 0.840 -0.009 0.100 Avg models 63.9 Ā± 0.0 147.0 Ā± 63.3 0.842 Ā± 0.011 0.004 Ā± 0.067 0.103 Ā± 0.015 it Model A 64.1 (79.1) 93.5 (42.8) 0.847 0.075 0.097 Model B 64.1 (78.3) 227.8 (76.8) 0.824 -0.065 0.069 Model C 64.1 (79.1) 91.6 (34.1) 0.847 0.072 0.094 Model D 64.1 (79.1) 139.2 (56.1) 0.838 0.016 0.084 Avg models 64.1 Ā± 0.0 138.0 Ā± 63.8 0.839 Ā± 0.011 0.025 Ā± 0.066 0.086 Ā± 0.013 Avg langs Model A 63.1 Ā± 10.5 98.7 Ā± 12.7 0.851 Ā± 0.003 0.053 Ā± 0.024 0.113 Ā± 0.016 Model B 63.1 Ā± 10.6 238.4 Ā± 24.4 0.827 Ā± 0.004 -0.098 Ā± 0.027 0.078 Ā± 0.013 Model C 63.1 Ā± 10.5 101.8 Ā± 11.9 0.849 Ā± 0.002 0.042 Ā± 0.027 0.108 Ā± 0.017 Model D 63.1 Ā± 10.5 137.4 Ā± 11.4 0.840 Ā± 0.003 -0.017 Ā± 0.030 0.096 Ā± 0.015 profile de Model A 47.9 (56.3) 94.8 (38.9) 0.847 0.053 0.094 Model B 47.8 (56.3) 207.4 (68.9) 0.821 -0.112 0.060 Model C 47.9 (56.3) 98.2 (32.2) 0.845 0.041 0.085 Model D 47.9 (56.3) 165.6 (62.2) 0.832 -0.044 0.072 Avg models 47.9 Ā± 0.0 141.5 Ā± 54.7 0.836 Ā± 0.012 -0.016 Ā± 0.077 0.078 Ā± 0.015 en Model A 62.4 (78.7) 122.7 (60.7) 0.852 -0.003 0.119 Model B 62.5 (78.9) 263.5 (99.9) 0.830 -0.154 0.081 Model C 62.4 (78.7) 120.7 (48.8) 0.850 -0.017 0.115 Model D 62.4 (78.7) 180.1 (70.2) 0.837 -0.105 0.091 Avg models 62.4 Ā± 0.0 171.7 Ā± 67.1 0.842 Ā± 0.011 -0.070 Ā± 0.072 0.101 Ā± 0.018 es Model A 77.5 (96.5) 129.8 (53.8) 0.851 0.041 0.130 Model B 77.5 (96.3) 272.9 (86.2) 0.828 -0.107 0.093 Model C 77.5 (96.5) 127.5 (52.8) 0.849 0.030 0.126 Model D 77.5 (96.5) 200.8 (80.7) 0.838 -0.043 0.107 Avg models 77.5 Ā± 0.0 182.7 Ā± 69.1 0.841 Ā± 0.011 -0.020 Ā± 0.069 0.114 Ā± 0.017 fr Model A 63.9 (77.2) 117.8 (47.1) 0.849 0.047 0.114 Model B 63.9 (77.0) 242.0 (76.5) 0.827 -0.094 0.081 Model C 63.9 (77.2) 119.5 (46.7) 0.848 0.037 0.108 Model D 63.9 (77.2) 192.6 (75.8) 0.835 -0.042 0.092 Avg models 63.9 Ā± 0.0 168.0 Ā± 60.4 0.840 Ā± 0.011 -0.013 Ā± 0.067 0.099 Ā± 0.015 it Model A 64.1 (79.1) 109.6 (47.7) 0.846 0.064 0.094 Model B 64.0 (78.8) 230.6 (72.8) 0.823 -0.072 0.067 Model C 64.1 (79.1) 98.7 (34.1) 0.846 0.067 0.093 Model D 64.0 (79.1) 185.6 (73.3) 0.833 -0.011 0.078 Avg models 64.0 Ā± 0.0 156.1 Ā± 62.9 0.837 Ā± 0.011 0.012 Ā± 0.067 0.083 Ā± 0.013 Avg langs Model A 63.1 Ā± 10.5 114.9 Ā± 13.5 0.849 Ā± 0.003 0.040 Ā± 0.026 0.110 Ā± 0.016 Model B 63.1 Ā± 10.5 243.3 Ā± 26.2 0.826 Ā± 0.004 -0.108 Ā± 0.030 0.076 Ā± 0.013 Model C 63.1 Ā± 10.5 112.9 Ā± 13.6 0.848 Ā± 0.002 0.032 Ā± 0.031 0.105 Ā± 0.017 Model D 63.1 Ā± 10.5 184.9 Ā± 13.3 0.835 Ā± 0.003 -0.049 Ā± 0.034 0.088 Ā± 0.014 behavior de Model A 47.9 (56.3) 90.7 (54.9) 0.847 0.050 0.091 Model B 47.7 (55.8) 148.6 (74.4) 0.828 -0.067 0.068 Model C 47.9 (56.3) 96.9 (37.9) 0.846 0.041 0.084 Model D 47.9 (56.3) 135.4 (97.0) 0.837 -0.013 0.078 Avg models 47.8 Ā± 0.1 117.9 Ā± 28.5 0.839 Ā± 0.009 0.003 Ā± 0.054 0.080 Ā± 0.010 en Model A 62.4 (78.7) 155.9 (83.1) 0.847 -0.038 0.109 Model B 62.3 (78.7) 241.8 (109.1) 0.831 -0.149 0.084 Model C 62.4 (78.7) 132.7 (63.3) 0.849 -0.027 0.112 Model D 62.4 (78.7) 212.7 (125.1) 0.836 -0.111 0.089 Avg models 62.4 Ā± 0.1 185.8 Ā± 50.2 0.841 Ā± 0.009 -0.081 Ā± 0.059 0.098 Ā± 0.014 es Model A 77.5 (96.5) 167.2 (80.5) 0.845 0.002 0.117 Model B 77.4 (96.5) 230.2 (94.4) 0.829 -0.097 0.098 Model C 77.5 (96.5) 138.9 (63.4) 0.848 0.021 0.124 Model D 77.5 (96.5) 260.2 (135.5) 0.834 -0.065 0.099 Avg models 77.5 Ā± 0.0 199.1 Ā± 55.8 0.839 Ā± 0.009 -0.035 Ā± 0.055 0.109 Ā± 0.013 Table 11: Full results (part 1). Rows grouped by setup and language. Avg Human Words varies slightly across models and setups due to rare cases where human comments were excluded from evaluation owing to missing generated counterparts (generation errors or hallucinations). Setup Lang Model Avg Human Words Avg Agent Words BERT ā rBERT ā ROUGE ā behavior fr Model A 63.9 (77.2) 133.6 (67.5) 0.846 0.024 0.107 Model B 63.7 (77.2) 199.8 (78.2) 0.829 -0.081 0.086 Model C 63.9 (77.2) 124.6 (54.2) 0.847 0.032 0.106 Model D 63.9 (77.2) 194.5 (115.4) 0.836 -0.037 0.092 Avg models 63.8 Ā± 0.1 163.1 Ā± 39.6 0.839 Ā± 0.008 -0.016 Ā± 0.054 0.098 Ā± 0.011 it Model A 64.1 (79.1) 130.8 (70.9) 0.842 0.042 0.090 Model B 63.9 (78.6) 183.4 (76.2) 0.826 -0.056 0.072 Model C 64.1 (79.1) 104.2 (41.7) 0.845 0.062 0.092 Model D 64.1 (79.1) 211.9 (126.5) 0.831 -0.022 0.076 Avg models 64.0 Ā± 0.1 157.6 Ā± 49.0 0.836 Ā± 0.009 0.006 Ā± 0.055 0.082 Ā± 0.010 Avg langs Model A 63.1 Ā± 10.5 135.6 Ā± 29.4 0.845 Ā± 0.002 0.016 Ā± 0.035 0.103 Ā± 0.012 Model B 63.0 Ā± 10.5 200.8 Ā± 37.3 0.829 Ā± 0.002 -0.090 Ā± 0.036 0.082 Ā± 0.012 Model C 63.1 Ā± 10.5 119.4 Ā± 18.2 0.847 Ā± 0.001 0.026 Ā± 0.033 0.104 Ā± 0.016 Model D 63.1 Ā± 10.5 203.0 Ā± 45.0 0.835 Ā± 0.002 -0.050 Ā± 0.040 0.087 Ā± 0.010 behavior+r de Model A 47.9 (56.3) 84.7 (50.8) 0.853 0.085 0.112 Model B 47.6 (56.2) 156.8 (76.8) 0.829 -0.058 0.072 Model C 47.9 (56.3) 95.8 (36.8) 0.847 0.050 0.089 Model D 47.9 (56.3) 140.0 (86.5) 0.839 -0.001 0.085 Avg models 47.8 Ā± 0.1 119.3 Ā± 34.6 0.842 Ā± 0.010 0.019 Ā± 0.062 0.089 Ā± 0.017 en Model A 62.4 (78.7) 127.0 (77.1) 0.854 0.013 0.131 Model B 62.4 (78.0) 243.8 (114.3) 0.833 -0.134 0.086 Model C 62.4 (78.7) 123.8 (55.8) 0.850 -0.015 0.116 Model D 62.4 (78.7) 190.2 (103.2) 0.840 -0.085 0.098 Avg models 62.4 Ā± 0.0 171.2 Ā± 57.3 0.844 Ā± 0.010 -0.055 Ā± 0.067 0.108 Ā± 0.020 es Model A 77.5 (96.5) 149.0 (74.5) 0.852 0.046 0.138 Model B 77.2 (96.6) 246.7 (95.2) 0.831 -0.087 0.099 Model C 77.5 (96.5) 131.6 (59.1) 0.850 0.035 0.130 Model D 77.5 (96.5) 241.6 (117.5) 0.838 -0.040 0.109 Avg models 77.4 Ā± 0.1 192.2 Ā± 60.4 0.843 Ā± 0.010 -0.011 Ā± 0.063 0.119 Ā± 0.018 fr Model A 63.9 (77.2) 124.7 (62.5) 0.850 0.051 0.120 Model B 63.6 (76.9) 216.2 (83.3) 0.830 -0.076 0.086 Model C 63.9 (77.2) 121.4 (51.9) 0.848 0.039 0.109 Model D 63.9 (77.2) 195.6 (100.4) 0.837 -0.027 0.097 Avg models 63.8 Ā± 0.1 164.5 Ā± 48.6 0.841 Ā± 0.009 -0.003 Ā± 0.060 0.103 Ā± 0.015 it Model A 64.1 (79.1) 117.5 (67.9) 0.848 0.079 0.110 Model B 63.9 (78.9) 194.4 (83.0) 0.827 -0.047 0.073 Model C 64.1 (79.1) 101.5 (40.6) 0.847 0.070 0.096 Model D 64.0 (79.1) 204.5 (113.4) 0.834 -0.004 0.083 Avg models 64.0 Ā± 0.1 154.5 Ā± 52.5 0.839 Ā± 0.010 0.024 Ā± 0.060 0.091 Ā± 0.016 Avg langs Model A 63.1 Ā± 10.5 120.6 Ā± 23.2 0.851 Ā± 0.003 0.055 Ā± 0.029 0.122 Ā± 0.012 Model B 62.9 Ā± 10.5 211.6 Ā± 37.4 0.830 Ā± 0.002 -0.081 Ā± 0.034 0.083 Ā± 0.011 Model C 63.1 Ā± 10.5 114.8 Ā± 15.4 0.848 Ā± 0.002 0.036 Ā± 0.031 0.108 Ā± 0.016 Model D 63.1 Ā± 10.5 194.4 Ā± 36.4 0.838 Ā± 0.002 -0.031 Ā± 0.034 0.094 Ā± 0.011 history de Model A 47.9 (56.3) 67.7 (38.3) 0.856 0.105 0.114 Model B 47.0 (55.2) 154.8 (65.0) 0.829 -0.064 0.070 Model C 47.9 (56.3) 85.3 (31.7) 0.848 0.059 0.091 Model D 47.9 (56.3) 130.7 (49.6) 0.837 -0.013 0.079 Avg models 47.6 Ā± 0.4 109.6 Ā± 40.1 0.842 Ā± 0.012 0.022 Ā± 0.075 0.089 Ā± 0.019 en Model A 62.4 (78.7) 96.7 (58.2) 0.860 0.054 0.145 Model B 62.2 (78.1) 203.8 (96.6) 0.837 -0.108 0.095 Model C 62.4 (78.7) 103.4 (47.6) 0.854 0.012 0.127 Model D 62.4 (78.7) 147.3 (62.9) 0.843 -0.063 0.105 Avg models 62.4 Ā± 0.1 137.8 Ā± 49.4 0.849 Ā± 0.011 -0.026 Ā± 0.073 0.118 Ā± 0.023 es Model A 77.5 (96.5) 103.1 (55.9) 0.859 0.091 0.154 Model B 77.3 (96.7) 210.8 (87.7) 0.836 -0.058 0.109 Model C 77.5 (96.5) 119.3 (51.9) 0.853 0.052 0.139 Model D 77.5 (96.5) 168.0 (65.7) 0.843 -0.008 0.119 Avg models 77.4 Ā± 0.1 150.3 Ā± 48.9 0.848 Ā± 0.010 0.019 Ā± 0.065 0.130 Ā± 0.020 fr Model A 63.9 (77.2) 90.0 (49.6) 0.858 0.104 0.142 Model B 63.4 (76.5) 191.7 (81.2) 0.834 -0.050 0.095 Model C 63.9 (77.2) 105.5 (45.8) 0.852 0.062 0.121 Model D 63.9 (77.2) 159.3 (62.6) 0.841 -0.004 0.107 Avg models 63.8 Ā± 0.2 136.6 Ā± 47.2 0.846 Ā± 0.011 0.028 Ā± 0.068 0.116 Ā± 0.020 it Model A 64.1 (79.1) 81.9 (48.5) 0.852 0.103 0.108 Model B 63.0 (78.4) 173.7 (71.7) 0.829 -0.034 0.075 Model C 64.1 (79.1) 91.5 (36.3) 0.848 0.077 0.098 Model D 64.1 (79.1) 155.2 (59.8) 0.837 0.014 0.085 Avg models 63.8 Ā± 0.5 125.6 Ā± 45.7 0.842 Ā± 0.010 0.040 Ā± 0.062 0.091 Ā± 0.015 Avg langs Model A 63.1 Ā± 10.5 87.9 Ā± 13.8 0.857 Ā± 0.003 0.091 Ā± 0.022 0.133 Ā± 0.020 Model B 62.6 Ā± 10.7 187.0 Ā± 22.8 0.833 Ā± 0.004 -0.063 Ā± 0.027 0.089 Ā± 0.016 Model C 63.1 Ā± 10.5 101.0 Ā± 13.2 0.851 Ā± 0.003 0.053 Ā± 0.024 0.115 Ā± 0.020 Model D 63.1 Ā± 10.5 152.1 Ā± 14.1 0.840 Ā± 0.003 -0.015 Ā± 0.029 0.099 Ā± 0.017 Table 12: Full results (part 2). Rows grouped by setup and language. Avg Human Words varies slightly across models and setups due to rare cases where human comments were excluded from evaluation owing to missing generated counterparts (generation errors or hallucinations). Sentiment Dialog Acts WD ā JS ā Density Ratio CRIT NEG POS PART INF REP COM REQ ASK REF base de Model A 0.236 0.147 0.700 0.002 0.051 -0.027 -0.009 0.042 0.040 0.005 -0.107 0.003 0.000 Model B 0.565 0.270 0.549 0.018 0.077 0.034 0.007 0.078 0.057 -0.017 -0.165 0.036 -0.123 Model C 0.329 0.259 0.733 0.028 0.073 -0.061 -0.021 0.069 0.081 -0.027 -0.151 -0.019 0.029 Model D 0.459 0.207 0.733 -0.010 0.035 -0.000 -0.000 -0.020 0.066 -0.028 -0.100 0.069 -0.013 Avg 0.397 Ā± 0.145 0.221 Ā± 0.056 0.679 Ā± 0.088 0.009 0.059 -0.014 -0.006 0.042 0.061 -0.017 -0.131 0.022 -0.027 en Model A 0.188 0.114 0.750 -0.001 0.042 -0.011 -0.020 0.022 0.048 -0.004 -0.056 0.001 -0.021 Model B 0.486 0.247 0.591 0.010 0.066 0.038 -0.001 0.058 0.067 -0.053 -0.138 0.044 -0.091 Model C 0.171 0.193 0.764 0.012 0.056 -0.030 -0.046 0.040 0.066 0.003 -0.128 0.009 0.017 Model D 0.415 0.198 0.715 -0.010 0.035 0.007 -0.007 -0.011 0.065 -0.034 -0.074 0.069 -0.040 Avg 0.315 Ā± 0.159 0.188 Ā± 0.055 0.705 Ā± 0.079 0.003 0.050 0.001 -0.019 0.027 0.062 -0.022 -0.099 0.031 -0.034 es Model A 0.339 0.178 0.700 0.007 0.078 -0.030 -0.020 0.048 0.063 0.004 -0.082 -0.031 -0.037 Model B 0.580 0.272 0.579 0.012 0.104 0.032 0.003 0.070 0.068 -0.005 -0.142 0.001 -0.141 Model C 0.371 0.298 0.771 0.023 0.095 -0.033 -0.072 0.087 0.082 0.022 -0.080 -0.158 0.035 Model D 0.517 0.176 0.724 -0.000 0.047 0.020 -0.010 -0.015 0.060 0.009 -0.076 0.038 -0.072 Avg 0.452 Ā± 0.115 0.231 Ā± 0.063 0.694 Ā± 0.082 0.010 0.081 -0.003 -0.025 0.048 0.068 0.007 -0.095 -0.038 -0.054 fr Model A 0.239 0.156 0.699 0.006 0.049 -0.028 -0.007 0.048 0.064 -0.017 -0.087 -0.015 -0.012 Model B 0.501 0.277 0.570 0.014 0.078 0.030 0.012 0.084 0.076 -0.037 -0.162 0.025 -0.120 Model C 0.252 0.263 0.743 0.034 0.076 -0.074 -0.048 0.081 0.084 -0.022 -0.108 -0.052 0.029 Model D 0.404 0.217 0.694 0.001 0.040 0.006 -0.005 0.001 0.074 -0.028 -0.101 0.059 -0.047 Avg 0.349 Ā± 0.126 0.228 Ā± 0.054 0.676 Ā± 0.074 0.014 0.061 -0.016 -0.012 0.053 0.074 -0.026 -0.115 0.004 -0.037 it Model A 0.219 0.155 0.710 -0.006 0.054 -0.055 0.002 0.043 0.053 -0.000 -0.093 0.017 -0.014 Model B 0.559 0.296 0.566 0.005 0.089 0.015 0.009 0.090 0.063 -0.015 -0.153 0.051 -0.154 Model C 0.205 0.258 0.768 0.018 0.083 -0.110 -0.048 0.074 0.067 0.009 -0.144 0.019 0.032 Model D 0.462 0.198 0.693 -0.010 0.039 0.001 0.006 -0.005 0.051 0.003 -0.088 0.064 -0.060 Avg 0.361 Ā± 0.177 0.227 Ā± 0.063 0.684 Ā± 0.085 0.002 0.066 -0.037 -0.008 0.050 0.058 -0.001 -0.120 0.038 -0.049 Avg Model A 0.244 Ā± 0.057 0.150 Ā± 0.023 0.712 Ā± 0.022 0.001 0.055 -0.030 -0.011 0.041 0.053 -0.003 -0.085 -0.005 -0.017 Model B 0.538 Ā± 0.042 0.272 Ā± 0.018 0.571 Ā± 0.015 0.012 0.083 0.030 0.006 0.076 0.066 -0.025 -0.152 0.031 -0.126 Model C 0.266 Ā± 0.084 0.254 Ā± 0.038 0.756 Ā± 0.017 0.023 0.077 -0.062 -0.047 0.070 0.076 -0.003 -0.122 -0.040 0.029 Model D 0.451 Ā± 0.045 0.199 Ā± 0.015 0.712 Ā± 0.018 -0.006 0.039 0.007 -0.003 -0.010 0.063 -0.016 -0.088 0.060 -0.046 profile de Model A 0.215 0.134 0.701 0.002 0.046 -0.020 -0.012 0.019 0.035 -0.009 -0.094 0.019 0.015 Model B 0.566 0.290 0.537 0.016 0.078 0.040 0.008 0.075 0.061 -0.033 -0.158 0.053 -0.141 Model C 0.384 0.263 0.721 0.029 0.073 -0.064 -0.021 0.056 0.081 -0.045 -0.145 0.000 0.035 Model D 0.452 0.215 0.688 0.005 0.045 -0.002 -0.012 -0.018 0.061 -0.058 -0.075 0.071 -0.017 Avg 0.404 Ā± 0.147 0.225 Ā± 0.069 0.662 Ā± 0.084 0.013 0.061 -0.012 -0.009 0.033 0.059 -0.036 -0.118 0.036 -0.027 en Model A 0.165 0.113 0.736 -0.004 0.041 -0.019 -0.019 0.014 0.049 -0.014 -0.052 0.015 -0.011 Model B 0.487 0.271 0.564 0.011 0.068 0.041 0.003 0.056 0.073 -0.062 -0.132 0.057 -0.114 Model C 0.223 0.212 0.741 0.017 0.060 -0.051 -0.045 0.029 0.079 -0.010 -0.125 0.018 0.026 Model D 0.294 0.212 0.655 0.008 0.045 -0.021 -0.011 -0.009 0.068 -0.062 -0.054 0.069 -0.033 Avg 0.292 Ā± 0.140 0.202 Ā± 0.065 0.674 Ā± 0.083 0.008 0.053 -0.012 -0.018 0.022 0.067 -0.037 -0.091 0.040 -0.033 es Model A 0.307 0.166 0.700 0.002 0.072 -0.030 -0.019 0.033 0.066 -0.002 -0.081 -0.010 -0.030 Model B 0.585 0.283 0.568 0.012 0.104 0.032 0.008 0.064 0.071 -0.014 -0.143 0.019 -0.155 Model C 0.372 0.299 0.759 0.027 0.097 -0.036 -0.074 0.071 0.086 0.017 -0.083 -0.143 0.038 Model D 0.467 0.186 0.684 0.009 0.062 0.002 -0.011 -0.016 0.059 -0.003 -0.062 0.040 -0.080 Avg 0.433 Ā± 0.121 0.233 Ā± 0.067 0.678 Ā± 0.080 0.013 0.084 -0.008 -0.024 0.038 0.071 -0.000 -0.092 -0.023 -0.057 fr Model A 0.202 0.155 0.696 -0.001 0.047 -0.029 -0.011 0.030 0.065 -0.028 -0.082 0.008 0.001 Model B 0.505 0.289 0.554 0.013 0.078 0.032 0.015 0.079 0.077 -0.045 -0.155 0.040 -0.134 Model C 0.247 0.264 0.739 0.035 0.077 -0.079 -0.047 0.068 0.085 -0.029 -0.104 -0.039 0.033 Model D 0.323 0.228 0.652 0.010 0.051 -0.002 -0.012 0.003 0.073 -0.057 -0.081 0.060 -0.044 Avg 0.319 Ā± 0.134 0.234 Ā± 0.058 0.660 Ā± 0.079 0.014 0.063 -0.020 -0.014 0.045 0.075 -0.040 -0.105 0.017 -0.036 it Model A 0.224 0.166 0.698 -0.007 0.054 -0.057 0.000 0.031 0.055 -0.009 -0.093 0.035 -0.010 Model B 0.560 0.308 0.556 0.001 0.089 0.022 0.016 0.086 0.064 -0.019 -0.148 0.057 -0.167 Model C 0.209 0.270 0.758 0.021 0.086 -0.114 -0.052 0.065 0.069 0.004 -0.144 0.029 0.037 Model D 0.406 0.191 0.659 -0.001 0.046 -0.007 0.003 -0.006 0.043 -0.010 -0.066 0.064 -0.066 Avg 0.350 Ā± 0.166 0.234 Ā± 0.066 0.668 Ā± 0.085 0.003 0.069 -0.039 -0.008 0.044 0.058 -0.009 -0.113 0.046 -0.052 Avg Model A 0.223 Ā± 0.052 0.147 Ā± 0.023 0.706 Ā± 0.017 -0.002 0.052 -0.031 -0.012 0.025 0.054 -0.012 -0.080 0.013 -0.007 Model B 0.541 Ā± 0.042 0.288 Ā± 0.013 0.556 Ā± 0.012 0.010 0.084 0.033 0.010 0.072 0.069 -0.035 -0.147 0.045 -0.142 Model C 0.287 Ā± 0.084 0.262 Ā± 0.031 0.744 Ā± 0.016 0.026 0.079 -0.069 -0.048 0.058 0.080 -0.013 -0.120 -0.027 0.034 Model D 0.389 Ā± 0.077 0.206 Ā± 0.017 0.668 Ā± 0.017 0.006 0.050 -0.006 -0.009 -0.009 0.061 -0.038 -0.068 0.061 -0.048 behavior de Model A 0.257 0.227 0.644 0.004 0.055 -0.012 -0.002 0.133 0.060 0.011 -0.142 -0.094 -0.013 Model B 0.551 0.246 0.493 0.010 0.073 0.035 0.023 0.141 -0.072 -0.003 -0.129 0.012 -0.090 Model C 0.360 0.281 0.687 0.030 0.074 -0.059 -0.021 0.112 0.084 -0.015 -0.177 -0.045 0.017 Model D 0.389 0.233 0.536 -0.002 0.059 0.009 0.007 0.114 0.038 -0.091 -0.125 0.046 -0.056 Avg 0.389 Ā± 0.122 0.247 Ā± 0.024 0.590 Ā± 0.091 0.011 0.065 -0.007 0.002 0.125 0.027 -0.024 -0.143 -0.020 -0.036 en Model A 0.278 0.181 0.632 -0.007 0.045 0.012 -0.006 0.087 0.070 0.004 -0.091 -0.049 -0.064 Model B 0.510 0.229 0.510 0.006 0.063 0.052 0.020 0.096 -0.025 -0.030 -0.098 0.038 -0.123 Model C 0.259 0.233 0.675 0.016 0.058 -0.023 -0.037 0.081 0.087 0.010 -0.166 0.000 -0.027 Model D 0.399 0.249 0.518 -0.004 0.052 0.024 0.008 0.089 0.054 -0.084 -0.091 0.063 -0.112 Avg 0.361 Ā± 0.117 0.223 Ā± 0.029 0.584 Ā± 0.083 0.003 0.055 0.016 -0.004 0.088 0.047 -0.025 -0.111 0.013 -0.081 es Model A 0.457 0.235 0.598 0.008 0.079 0.010 -0.015 0.100 0.073 0.007 -0.113 -0.060 -0.090 Model B 0.584 0.241 0.536 -0.000 0.096 0.040 0.028 0.094 -0.012 -0.009 -0.102 0.012 -0.146 Model C 0.413 0.299 0.713 0.028 0.097 -0.020 -0.065 0.097 0.086 0.021 -0.112 -0.137 0.005 Model D 0.536 0.234 0.529 -0.000 0.080 0.036 0.009 0.088 0.052 -0.032 -0.119 0.021 -0.134 Avg 0.498 Ā± 0.077 0.253 Ā± 0.031 0.594 Ā± 0.085 0.009 0.088 0.017 -0.011 0.095 0.050 -0.003 -0.112 -0.041 -0.091 Table 13: Dialog acts (part 1). JS = Jensen-Shannon distance on label proportions. Density Ratio = human labels/sentence Ć· agent labels/sentence. Dialog Acts = human proportion ā- agent proportion for each label (positive ā agent under-produces). Sentiment Dialog Acts WD ā JS ā Density Ratio CRIT NEG POS PART INF REP COM REQ ASK REF behavior fr Model A 0.358 0.222 0.615 0.007 0.055 -0.004 -0.001 0.115 0.075 -0.007 -0.120 -0.069 -0.051 Model B 0.508 0.246 0.511 0.007 0.073 0.035 0.030 0.125 -0.011 -0.040 -0.124 0.025 -0.120 Model C 0.323 0.280 0.689 0.038 0.076 -0.057 -0.050 0.106 0.088 -0.012 -0.127 -0.070 0.009 Model D 0.457 0.291 0.524 0.004 0.064 0.034 0.008 0.114 0.068 -0.108 -0.133 0.054 -0.106 Avg 0.411 Ā± 0.086 0.260 Ā± 0.031 0.585 Ā± 0.084 0.014 0.067 0.002 -0.003 0.115 0.055 -0.042 -0.126 -0.015 -0.067 it Model A 0.358 0.204 0.609 -0.006 0.061 -0.019 0.006 0.112 0.060 -0.007 -0.123 -0.019 -0.064 Model B 0.556 0.266 0.512 -0.009 0.084 0.039 0.037 0.127 -0.028 -0.028 -0.113 0.036 -0.146 Model C 0.285 0.277 0.717 0.021 0.086 -0.095 -0.053 0.102 0.072 0.012 -0.169 0.013 0.012 Model D 0.484 0.259 0.501 -0.010 0.069 0.026 0.021 0.105 0.030 -0.065 -0.109 0.059 -0.125 Avg 0.421 Ā± 0.122 0.252 Ā± 0.032 0.585 Ā± 0.100 -0.001 0.075 -0.012 0.003 0.112 0.034 -0.022 -0.129 0.022 -0.081 Avg Model A 0.342 Ā± 0.079 0.214 Ā± 0.022 0.620 Ā± 0.018 0.001 0.059 -0.003 -0.003 0.109 0.068 0.002 -0.118 -0.058 -0.056 Model B 0.542 Ā± 0.033 0.245 Ā± 0.014 0.512 Ā± 0.016 0.003 0.078 0.040 0.028 0.117 -0.030 -0.022 -0.113 0.025 -0.125 Model C 0.328 Ā± 0.061 0.274 Ā± 0.025 0.696 Ā± 0.018 0.027 0.078 -0.051 -0.045 0.099 0.083 0.003 -0.150 -0.048 0.003 Model D 0.453 Ā± 0.061 0.253 Ā± 0.024 0.522 Ā± 0.013 -0.002 0.065 0.026 0.011 0.102 0.048 -0.076 -0.115 0.049 -0.107 behavior+r de Model A 0.155 0.160 0.701 -0.002 0.042 -0.021 -0.002 0.084 0.032 0.004 -0.096 -0.069 0.027 Model B 0.534 0.225 0.524 0.004 0.070 0.025 0.021 0.122 -0.002 -0.030 -0.134 0.010 -0.086 Model C 0.343 0.266 0.686 0.025 0.067 -0.066 -0.021 0.110 0.079 -0.020 -0.164 -0.050 0.038 Model D 0.289 0.150 0.583 0.004 0.044 -0.014 0.005 0.084 -0.014 -0.038 -0.097 0.030 -0.004 Avg 0.330 Ā± 0.157 0.200 Ā± 0.055 0.623 Ā± 0.085 0.008 0.056 -0.019 0.001 0.100 0.024 -0.021 -0.123 -0.020 -0.006 en Model A 0.168 0.111 0.705 -0.005 0.032 -0.004 -0.004 0.059 0.041 -0.003 -0.056 -0.040 -0.019 Model B 0.477 0.228 0.532 0.004 0.056 0.038 0.017 0.094 0.037 -0.060 -0.109 0.038 -0.114 Model C 0.191 0.208 0.686 0.015 0.049 -0.042 -0.030 0.086 0.083 0.006 -0.142 -0.017 -0.007 Model D 0.265 0.151 0.565 0.002 0.036 -0.008 0.003 0.070 0.008 -0.048 -0.061 0.046 -0.048 Avg 0.275 Ā± 0.141 0.175 Ā± 0.053 0.622 Ā± 0.086 0.004 0.043 -0.004 -0.004 0.077 0.042 -0.026 -0.092 0.007 -0.047 es Model A 0.338 0.157 0.673 0.004 0.059 -0.015 -0.004 0.063 0.054 -0.000 -0.086 -0.041 -0.034 Model B 0.585 0.238 0.562 0.002 0.083 0.032 0.023 0.088 0.047 -0.032 -0.121 0.017 -0.138 Model C 0.349 0.275 0.721 0.022 0.081 -0.037 -0.050 0.100 0.084 0.019 -0.110 -0.131 0.022 Model D 0.449 0.161 0.564 0.003 0.058 0.014 0.007 0.069 0.018 -0.017 -0.085 0.021 -0.088 Avg 0.430 Ā± 0.114 0.208 Ā± 0.058 0.630 Ā± 0.080 0.008 0.070 -0.001 -0.006 0.080 0.050 -0.008 -0.100 -0.034 -0.059 fr Model A 0.240 0.161 0.671 0.001 0.043 -0.021 -0.000 0.076 0.055 -0.012 -0.090 -0.058 0.006 Model B 0.499 0.249 0.535 0.007 0.067 0.029 0.029 0.110 0.047 -0.057 -0.139 0.021 -0.113 Model C 0.242 0.268 0.699 0.033 0.069 -0.062 -0.043 0.107 0.085 -0.016 -0.119 -0.084 0.031 Model D 0.353 0.188 0.556 0.008 0.048 0.004 0.005 0.086 0.024 -0.056 -0.112 0.040 -0.046 Avg 0.333 Ā± 0.122 0.216 Ā± 0.050 0.615 Ā± 0.082 0.012 0.057 -0.013 -0.002 0.095 0.053 -0.035 -0.115 -0.020 -0.031 it Model A 0.229 0.142 0.685 -0.008 0.046 -0.034 0.006 0.067 0.046 -0.006 -0.097 -0.016 -0.004 Model B 0.542 0.267 0.542 -0.013 0.078 0.037 0.029 0.116 0.037 -0.051 -0.130 0.039 -0.142 Model C 0.194 0.255 0.722 0.017 0.074 -0.097 -0.040 0.106 0.069 0.007 -0.159 -0.012 0.034 Model D 0.391 0.180 0.528 -0.007 0.054 0.005 0.014 0.086 -0.006 -0.035 -0.084 0.046 -0.073 Avg 0.339 Ā± 0.160 0.211 Ā± 0.060 0.619 Ā± 0.099 -0.003 0.063 -0.022 0.002 0.094 0.037 -0.021 -0.118 0.014 -0.046 Avg Model A 0.226 Ā± 0.073 0.146 Ā± 0.021 0.687 Ā± 0.016 -0.002 0.044 -0.019 -0.001 0.070 0.046 -0.003 -0.085 -0.045 -0.005 Model B 0.527 Ā± 0.042 0.241 Ā± 0.017 0.539 Ā± 0.014 0.001 0.071 0.032 0.024 0.106 0.033 -0.046 -0.127 0.025 -0.119 Model C 0.264 Ā± 0.078 0.254 Ā± 0.027 0.703 Ā± 0.018 0.023 0.068 -0.061 -0.037 0.102 0.080 -0.001 -0.139 -0.059 0.024 Model D 0.349 Ā± 0.075 0.166 Ā± 0.017 0.559 Ā± 0.020 0.002 0.048 0.000 0.007 0.079 0.006 -0.039 -0.088 0.037 -0.052 history de Model A 0.136 0.124 0.820 0.011 0.030 -0.024 -0.012 0.005 -0.003 -0.004 -0.074 0.028 0.044 Model B 0.558 0.261 0.592 0.019 0.075 0.031 0.004 0.076 0.057 -0.036 -0.151 0.037 -0.112 Model C 0.316 0.260 0.791 0.026 0.062 -0.027 -0.042 0.027 0.078 -0.036 -0.150 -0.001 0.063 Model D 0.416 0.204 0.736 -0.007 0.022 0.003 -0.009 -0.030 0.067 -0.040 -0.081 0.070 0.005 Avg 0.356 Ā± 0.177 0.212 Ā± 0.065 0.735 Ā± 0.101 0.012 0.047 -0.004 -0.015 0.019 0.050 -0.029 -0.114 0.034 0.000 en Model A 0.140 0.085 0.816 -0.002 0.022 -0.011 -0.013 -0.001 0.027 -0.011 -0.049 0.023 0.016 Model B 0.467 0.244 0.615 0.021 0.064 0.019 -0.002 0.062 0.061 -0.076 -0.122 0.051 -0.078 Model C 0.097 0.200 0.819 0.008 0.044 -0.004 -0.062 0.008 0.058 0.002 -0.123 0.011 0.057 Model D 0.375 0.190 0.705 -0.006 0.025 -0.004 -0.013 -0.014 0.065 -0.045 -0.061 0.069 -0.016 Avg 0.270 Ā± 0.180 0.180 Ā± 0.068 0.739 Ā± 0.098 0.005 0.039 0.000 -0.022 0.014 0.053 -0.033 -0.089 0.038 -0.005 es Model A 0.237 0.127 0.814 0.010 0.048 -0.051 -0.015 0.007 0.046 -0.005 -0.072 0.016 0.017 Model B 0.571 0.254 0.623 0.020 0.098 0.025 -0.000 0.072 0.060 -0.019 -0.135 0.012 -0.131 Model C 0.323 0.276 0.801 0.017 0.077 0.005 -0.091 0.047 0.079 0.018 -0.086 -0.123 0.057 Model D 0.492 0.159 0.735 0.002 0.030 0.016 -0.019 -0.028 0.058 0.004 -0.057 0.041 -0.046 Avg 0.406 Ā± 0.153 0.204 Ā± 0.072 0.743 Ā± 0.087 0.012 0.063 -0.001 -0.031 0.025 0.061 -0.001 -0.087 -0.014 -0.026 fr Model A 0.164 0.125 0.795 0.004 0.030 -0.039 -0.008 0.016 0.041 -0.022 -0.072 0.019 0.032 Model B 0.492 0.262 0.604 0.021 0.076 0.017 0.013 0.085 0.065 -0.053 -0.150 0.030 -0.104 Model C 0.201 0.263 0.782 0.028 0.065 -0.044 -0.068 0.049 0.080 -0.034 -0.101 -0.043 0.069 Model D 0.379 0.195 0.711 0.002 0.022 0.007 -0.011 -0.017 0.066 -0.038 -0.081 0.059 -0.010 Avg 0.309 Ā± 0.154 0.211 Ā± 0.066 0.723 Ā± 0.088 0.014 0.048 -0.015 -0.019 0.033 0.063 -0.037 -0.101 0.016 -0.003 it Model A 0.152 0.137 0.814 -0.001 0.028 -0.057 -0.002 0.006 0.042 -0.003 -0.080 0.035 0.030 Model B 0.550 0.283 0.599 0.013 0.087 0.007 0.004 0.092 0.059 -0.031 -0.143 0.050 -0.138 Model C 0.166 0.263 0.808 0.017 0.072 -0.062 -0.084 0.045 0.067 0.010 -0.150 0.027 0.058 Model D 0.448 0.182 0.700 -0.012 0.020 0.004 -0.001 -0.019 0.053 -0.004 -0.078 0.064 -0.027 Avg 0.329 Ā± 0.201 0.217 Ā± 0.069 0.730 Ā± 0.102 0.004 0.052 -0.027 -0.021 0.031 0.055 -0.007 -0.113 0.044 -0.019 Avg Model A 0.166 Ā± 0.041 0.120 Ā± 0.020 0.812 Ā± 0.009 0.004 0.031 -0.037 -0.010 0.007 0.031 -0.009 -0.069 0.024 0.028 Model B 0.527 Ā± 0.045 0.261 Ā± 0.014 0.607 Ā± 0.013 0.019 0.080 0.020 0.004 0.077 0.060 -0.043 -0.140 0.036 -0.112 Model C 0.221 Ā± 0.098 0.253 Ā± 0.030 0.800 Ā± 0.015 0.019 0.064 -0.026 -0.069 0.035 0.072 -0.008 -0.122 -0.026 0.061 Model D 0.422 Ā± 0.049 0.186 Ā± 0.017 0.717 Ā± 0.017 -0.004 0.024 0.005 -0.011 -0.022 0.062 -0.025 -0.072 0.061 -0.019 Table 14: Dialog acts (overall). JS = Jensen-Shannon distance on label proportions. Density Ratio = human labels/sentence Ć· agent labels/sentence. Dialog Acts = human proportion ā- agent proportion for each label (positive ā agent under-produces). Appendix H Information Density vs. Information Type A natural concern with our experimental design is that the comparison across conditioning methods may not be fully information-balanced: the history method receives 5ā20 raw prior comments, whereas profile and behavior receive compressed summaries. Part of the gain observed under history could therefore reflect a larger context budget (i.e., information quantity) rather than a strictly better form of persona representation, i.e., information type. To probe this confound, we examine how performance varies with the amount of conditioning information provided. We focus this analysis on profile length, where the available token budget varies meaningfully across users (see Table 2 and Figure 7). For history, the analogous analysis would require comparing performance across different numbers of prior comments. However, the history sampler defaults to using up to 20 comments per user, and in practice the vast majority of conditioning pools fall in the 16ā20 range; lower-comment buckets are too sparsely populated to support a meaningful comparison. We therefore restrict the analysis to profile length, which provides a well-distributed and interpretable proxy for input richness. We stratify all experimental instances by profile word count into five buckets (1ā50, 51ā100, 101ā150, 151ā200, 201ā250) and compute, for each bucket and for each languageāmodel pair, the four metrics used in our main analysis: ROUGE-L, rBERT, dialog act JSD, and sentiment JSD. Buckets with fewer than 20 samples are omitted, so de has no 201ā250 bucket and es no 151ā200 bucket. The es 201ā250 bucket is dominated by a single user, who contributes 45 of its 67 samples, so its values should be read as largely individual rather than bucket-level. Results are shown in Figures 20ā23. The red dashed lines indicate the proportion of samples falling in each bucket; trends at the longest bucket should be interpreted with caution due to sample sparsity. Findings Across all 20 languageāmodel pairs, none of the four metrics improves consistently as profile length grows. A pure context-budget account would predict a downward slope for JSD metrics (lower divergence = better alignment) and an upward slope for similarity metrics (higher = better) as more conditioning tokens are provided. ROUGE-L and rBERT are nearly horizontal in every configuration (within-pair ranges of at most 0.0260.026 and 0.0440.044, respectively). Where the JSD metrics do move, they mostly move in the opposite direction: dialog act JSD rises for all four models in fr, mainly from the 151ā200 bucket onward, and sentiment JSD rises with profile length for all four models in es, fr, and it. The few downward trends are small and confined to en and de: dialog act JSD for Model A in en decreases from 0.1350.135 to 0.0980.098, and sentiment JSD decreases for some models in en and de, in en largely because of the sparse 1ā50 bucket (under 3% of samples). Taken together, these results provide evidence against an information-density explanation for our main results. Together with the observation reported in the main textāthat profile degrades performance relative to base despite adding tokensāthis analysis supports our interpretation that information type, not information quantity, drives the differences across conditioning methods. We acknowledge that this analysis does not fully isolate every possible confound (e.g., compression-induced loss of stylistic exemplars), and we discuss this further in the Limitations. Figure 20: ROUGE-L between agent and human comments under profile conditioning, stratified by profile word count. Each row is a language; each column is a model. Brown solid line: ROUGE-L (left axis). Red dashed line: proportion of samples in each bucket (right axis). Figure 21: Rescaled BERTScore (rBERT) between agent and human comments under profile conditioning, stratified by profile word count. Conventions as in Figure 20. Figure 22: Dialog act JSD between agent and human distributions under profile conditioning, stratified by profile word count. Lower is better. Conventions as in Figure 20. Figure 23: Sentiment JSD between agent and human distributions under profile conditioning, stratified by profile word count. Lower is better. Conventions as in Figure 20. Appendix I Qualitative Audit of Profiles and Behavioral Traits Because profiles and behavioral traits are produced automatically by an LLM and then used as the key conditioning signals for profile, behavior, and behavior+r, their quality directly affects the interpretation of downstream results. To complement the quantitative analyses in the main text, we conducted a manual audit of these representations along three dimensions: (i) relevance and faithfulness of generated profiles with respect to the source user pages, (i) coherence between assigned behavioral traits and the supporting reasoning produced by the extractor model, and (i) privacy preservation, i.e., whether sensitive information (real names, usernames, contacts) is leaked into the generated representations. We sampled 10 users per language (50 total) uniformly at random from the experimental subset and one of the authors inspected, for each user, the generated profile, all 8 behavioral traits with their reasoning, and the underlying user page. Profile faithfulness In all 50 cases, profiles contain information that is relevant to and supported by the source user page. We identified a small number of imprecisions, none of which substantially distort the persona: ā¢ English: one numerical drift (ā42 yearsā instead of ā45 yearsā); one rounding error on edit count (ā100,000ā instead of ā100,000kā); ā¢ French: one case where contributed-article information was omitted; ā¢ German: one case where the username was incorporated into the profile as if it were a separate entity, likely due to hallucination triggered by user page ambiguity; ā¢ Italian: one ungrammatical fragment (āHa adottato il giorno 16 luglio e il comune di Plesioā) reflecting parsing ambiguity in the user page. We additionally observed that information from userboxesāsmall standardized templates that Wikipedia editors place on their user pages to declare interests, languages spoken, locations, or affiliationsāis sometimes incorporated into the profile and sometimes ignored, with no clear pattern across languages. Moreover, we noticed that gender, when not explicitly self-declared, is typically inferred from gender-marked terms (especially in the four grammatical-gender languages: German, Spanish, French, and Italian) and, when available, from given names. Behavioral trait coherence For each of the 8 behavioral categories, we checked whether the assigned trait value is consistent with the accompanying reasoning. Out of 50Ć8=40050Ć 8=400 trait-reasoning pairs, in only 7 cases (across English, Spanish, and Italian) was the assigned label found to trace back to the provided reasoning. In most other cases, the reasoning explicitly cites concrete excerpts from the userās comment history in support of the assigned trait, providing transparent grounding for the conditioning signal. Privacy The audit surfaced two cases that warrant explicit mention: ā¢ one English profile mentioning a middle name, which the user themselves had self-disclosed anecdotally on their user page; ā¢ one German profile reproducing the username as if it were a separate named entity (the same hallucination case noted above). No real legal names, contact details, or other directly identifying information were leaked into the generated representations. Implications for downstream results The audit indicates that the generated profiles and behavioral traits are largely faithful and well-grounded, with imprecisions concentrated at the level of isolated factual details rather than systematic distortions. This argues against an interpretation in which the relatively weak performance of profile and behavior is driven primarily by representation quality. At the same time, the audit highlights that profiles vary substantially in the kinds of information they emphasize (e.g., uneven uptake of userbox content) and that the compression step inherent in summarization may lose stylistic exemplars present in the raw history signal. In this sense, our results should be read as conclusions about the conditioning representations as we generated them, rather than as universal statements about all possible sociodemographic descriptors. We acknowledge this scope explicitly in the Limitations.