From Caricature to Collective Artifact: How Persona Construction Drives Polarization in Generative Social Simulations

• Generative social simulation uses LLM agents with textual personas to simulate social behavior, including interactions on synthetic social media platforms.Personas often derive from survey data and specify demographics, beliefs, political leanings, and interests.Source: §1 · p. 1
• Prior work documents persona biases and caricatures at the individual level, while how these representations shape collective outcomes remains less understood.Reported problems include stereotypical language, homogeneous output, limited emotional diversity, and stronger partisanship than the people represented.Source: §1 · p. 1
• Surveys record positions without necessarily measuring how much respondents care, but persona construction makes each answer an always-active, equally weighted description.The authors offer salience as one possible explanation, since LLMs tend to produce text consistent with their input.Source: §1 · p. 2
• The study tests whether personas amplify individual political attitudes, whether those distortions become network artifacts, and whether non-political preferences reduce polarization.It compares responses with ANES and conducts controlled simulation ablations while holding the platform fixed.Source: §1 · p. 2
• Full personas produce extremely polarized networks, while ablated agents form more cross-party ties, approaching empirical studies, even without losing all political orientation.No ablation condition matches all empirical benchmarks, and enriching personas with non-political statements does not reduce polarization.Source: §1 · p. 2
• Internal-persona ablation significantly affects collective outcomes, whereas removing partisan information from public biographies has little impact.Ablated agents also engage with less polarizing content.Source: §1 · p. 2

2 Related Work

• Reviews cover LLM-based social simulation broadly, while this paper focuses specifically on simulated social media interactions [30].Source: §2 · p. 2

2.1 Generative Social Media Simulation

• Related systems model social behavior with LLM agents, including memory and planning, platform-level information and emotion propagation [15], and large-scale opinion dynamics [49].Source: §2.1 · p. 3
• Generative social media simulations model agents that post, repost, and follow; this paper builds on a Twitter-like framework for news consumption and sharing [44] [23].Earlier systems also explore feed designs and network polarization, segregation, and hierarchy [23] [39].Source: §2.1 · p. 3
• Across varied systems, textual personas commonly drive agent behavior; this study instead examines sensitivity to persona specifications.Source: §2.1 · p. 3

2.2 Evaluating Personas for social simulation

• This study shifts attention from individual persona biases to their effects when caricatured personas are embedded in dynamic social network simulations.Source: §2.2 · p. 3
• Unlike one-shot network generation, this work embeds personas in dynamic social simulations and identifies internal persona instructions as a primary driver of polarized networks.Chang et al. found political affiliation to be the strongest homophily predictor in their one-shot setting.Source: §2.2 · p. 3
• Prior methodological work calls for mechanism ablations and shared evaluation protocols, while small writing-style changes have been shown to alter macro-level interactions [26] [50].The authors also distinguish their survey-derived personas from prior work using users’ previous posts on X [32].Source: §2.2 · p. 4

• LLMs infer the relative weight of persona statements, potentially giving partisan statements more influence than scalar opinion parameters do in traditional agent-based models.In traditional models, scalar values encode magnitude and direction, and modelers define their weight in update rules.Source: §3 · p. 4
• Survey-derived persona statements turn answers into always-active descriptions, although surveys do not measure how important each answer is to the respondent.The ablation changes which attributes are stated while keeping the underlying infrastructure fixed.Source: §3 · p. 4

4.1 Persona Construction

• Personas are sampled from the 2020 ANES and include demographic, political, ideological, issue, and interest information, plus separate internal prompts and public biographies.Internal personas guide agents’ actions; public biographies are shown to other agents when they consider reposting or following.Source: §4.1 · p. 4

4.2 Persona Ablation Design

• The cumulative ablations remove affective-polarization statements, party identification, and voting behavior while retaining demographics, ideology, important issues, and interests.Ablations are tested first in public biographies and then in both biographies and internal prompts; the full-persona condition retains all statements.Source: §4.2 · p. 5

4.3 Simulation Platform

• Agents read ten-post feeds, then may repost, share a headline with commentary, or take no action; reposting exposes the author’s biography and follower count.The feed algorithm, action space, news database, network initialization, and scheduling remain fixed while persona text changes.Source: §4.3 · pp. 5–6

4.4 Evaluation

• Before interacting, agents answer a three-part polarization questionnaire; separate API calls, repeated full personas, and standardized outputs support measurable comparisons with ANES responses.The questionnaire covers party feelings, trait attributions, and following behaviors.Source: §4.4 · p. 6

4.5 Comparison with Empirical Networks

• Two Twitter network studies provide external benchmarks, but neither represents ordinary users and both reflect specific political periods, so they are not treated as validation.One samples users who follow politicians; the other studies elected officials.Source: §4.5 · p. 6

4.6 Experimental Setup

• The experiments simulate 500 agents for 5,000 steps with 10 seeds per setting, using four language models; GPT-5.4-mini uses 200 personas and 2,000 steps because of cost.The study focuses on persona-design choices and robustness across models, not technical model performance.Source: §4.6 · pp. 6–7

4.7 Metrics

• Questionnaire metrics capture thermometer, trait, and follow gaps, while network metrics include the E-I index, clustering, and retweet-partisanship correlation.The thermometer gap compares mean own-party and opposite-party warmth; the E-I index compares external and internal links, with −1 indicating complete segregation.Source: §4.7 · p. 7
• The study also relates absolute partisan strength to received retweets and averages local clustering coefficients across nodes.Positive follow gaps indicate co-partisan preference; the trait gap ignores “don’t know” responses and separates positive from negative trait sets.Source: §4.7 · pp. 7–8

5.1 RQ1: Persona ablations alter individual political responses before interaction

• Removing AP alone brings agents’ thermometer gaps to ANES values; further ablations push them below the observed benchmark, though not to neutrality.Full personas exaggerate party and candidate thermometer polarization.Source: §5.1 · p. 9
• Under full personas, agents assign every positive trait to their own party and every negative trait to the opposing party; trait gaps shift only after PID removal.AP removal changes thermometer gaps but does not affect these personality-trait gaps.Source: §5.1 · p. 9
• Under full ablation, 46.0% of Democratic and 22.4% of Republican agents would follow opponents, versus 33% and 20% empirically; further removals can overshoot and leave Republicans more reluctant.The Republican–Democratic reluctance gap widens as more persona statements are removed.Source: §5.1 · p. 9

5.2 RQ2: Internal persona specifications drive collective polarization

• Public-biography ablations leave the E-I index negative, show no significant clustering change, and yield high variance in retweet-partisanship correlation.The average clustering coefficient trends downward, but the change is not statistically significant.Source: §5.2 · p. 10
• Internal ablation weakens the association between extreme partisan content and retweets; clustering shows no statistically significant change.Under full personas, extreme content from either party receives more retweets.Source: §5.2 · p. 10
• Removing AP and PID from internal prompts produces the largest significant shift from complete segregation toward almost full integration.The shift is reported at p < 0.05; public-biography ablations do not produce the same network change.Source: §5.2 · p. 10, Figure 4 · p. 11
• Ablated personas approach Twitter benchmarks (21.5% legislator ties; 33% liberal and 20% conservative politician-followers), but no condition matches all.For Republicans, removing AP and PID is sufficient to approach the legislator benchmark; Democrats also require removing voting behavior.Source: No configuration matches all empirical… · pp. 11–12, Figure 5 · p. 12

5.3 RQ3: Adding non-political statements does not reduce political polarization

• Adding soccer statements leaves political polarization unchanged, while soccer-based polarization increases as political statements are removed, plateaus quickly, and remains lower.The increase is not monotonic.Source: §5.3 · pp. 12–13, Figure 6 · p. 13

6 A Checklist for Persona Reporting in Generative Social Simulation

• Publish persona prompts verbatim, including source-data mappings and the rules and thresholds that generate statements; report internal personas separately from public-facing biographies.The two artifacts serve distinct functions and should not be treated as interchangeable.Source: §6 · p. 13
• State which persona attributes act as behavioral instructions or are intended to remain latent, then test marginal and cumulative effects through mechanism ablations across seeds.Also vary persona formulations while holding data and environment settings fixed.Source: §6 · p. 13
• Benchmark several empirical findings, report model sensitivity, and vary the component implicated by the claim while holding other components fixed.Matching one observation alone is not enough for behavioral validity.Source: §6 · pp. 13–14

7 Conclusion

• Removing polarizing statements from personas reduces collective polarization, while removing them only from public biographies has little effect.Adding non-political statements also does not reduce partisan identity’s influence on collective behavior.Source: §7 · p. 14
• Survey responses omit issue salience, but persona construction turns each answer into an always-active behavioral instruction.The authors suggest interview-based personas may reduce this problem because people voice what matters to them.Source: §7 · p. 14
• Personas are not the only possible source of distortion, and ablating them is not sufficient to establish valid collective dynamics.Interactions, communications, and the LLM itself may also contribute; the study did not explore how these mechanisms interact.Source: §7 · p. 14
• Future studies should sensitivity-test and benchmark persona construction, interaction, and communication protocols.Source: §7 · p. 14