Diverse Minds, Divided Networks? Personality Composition, Polarization, and Collective Intelligence in LLM-Based Social Simulations

2 Related work

• Earlier studies examined polarization or collective intelligence separately, so whether personality composition shapes both outcomes—and whether they trade off—remained unanswered.Trait manipulation and group-task measurement were absent from polarization studies, while collective-intelligence studies used small teams without social networks.Source: §1 · p. 2

1 Introduction

• Trait heterogeneity has the largest measured effects: more varied societies have more dispersed opinions but less segregation, while Agreeableness moderates Openness’s effect.The Agreeableness–Openness interaction replicated in a second model family; no polarization measure predicts poorer collective performance.Source: §1 · p. 2

2.1 Generative agent societies

• TraitMix uses a conventional social-media simulation environment; its contribution is controlling personality composition and measuring both outcome families in the same runs.The environment includes an interest- and popularity-weighted feed, bounded memory, follow actions, and free-text posting.Source: §1 · p. 2, §2.1 · p. 5

2.2 Opinion dynamics with language-model agents

• Prior LLM opinion-dynamics work found convergence and echo chambers, but typically treated agents as uniform and did not systematically vary personality or measure collective task performance.Cau et al. identified absent personalities and network structure as limitations; TraitMix manipulates composition and measures network segregation.Source: §2.2 · p. 5

2.3 Personality in language models

• LLM personality research supplies a caution: questionnaire-based induction may not persist through extended interaction, so questionnaire validation alone does not establish behavioral validity.The paper explicitly limits its own induction gate to questionnaire measurement.Source: §2.3 · p. 6

2.4 Collective intelligence and group composition

• Prior personality studies of collective intelligence used small teams without social networks and did not measure polarization.Without an opinion to polarize, those studies could not assess both outcomes together.Source: §2.4 · p. 6

2.5 Validity in generative social simulation

• In response to validation concerns, the study uses an induction gate, a neutral filler topic, reconstructed trait distributions, internal replications, robustness audits, and a second-model-family replication.The authors report measures that failed as well as those that passed.Source: §2.5 · p. 6

2.6 What is missing

• The central gap is that prior work neither manipulated personality composition while measuring both outcome families in the same societies nor made their relationship observable.Separating dispersion from segregation also makes it possible to test whether those two aspects of polarization move in opposite directions.Source: §2.6 · pp. 7–8

3 The TraitMix design

• TraitMix varies Big Five composition while measuring polarization and collective intelligence in the same simulation run.The design formalizes the framework, agent architecture, induction procedure, outcome batteries, and artifact controls.Source: §3 · p. 8

3.1 Problem formulation

• Personality composition is the distribution of Big Five vectors: its mean specifies trait levels and its covariance specifies heterogeneity.Experimental conditions choose the distribution’s mean and covariance while holding other simulation aspects fixed; realized moments are reconstructed because truncation can alter dispersion.Source: §3.1 · p. 8

3.2 Agent architecture and interaction

• Agents interact through a dynamic network and recommender-driven feed, while private probes measure beliefs rather than willingness to post.Follow and unfollow actions modify the network; opinions and collective tasks are elicited outside the platform.Source: §3.2 · p. 9

3.3 Personality induction and validation

• Trait prompts render each Big Five dimension at one of nine levels, and induced personality is checked by questionnaire and by scoring public posts.The induction gate is applied separately to each model before experimental runs; its magnitude threshold was added after observing a case where rank agreement overstated separation.Source: §3.3 · pp. 9–10

3.4 Outcome batteries

• Both outcome families are measured on the same societies in the same runs, making their relationship observable within one system.The task families are elicited privately before and after a discussion window.Source: §3.4 · p. 10, §3.4.2 · p. 11

3.4.1 Polarization

• Polarization is divided into opinion dispersion and segregation, which can move in opposite directions rather than form one interchangeable measure.When opinion variance collapses, echo-chamber closure can rise mechanically, so closure is interpreted alongside variance.Source: §3.4.1 · p. 10

3.4.2 Collective intelligence

• Collective intelligence is measured through estimation and hidden-profile decision tasks with verifiable answers, embedded in the discussion and elicited privately.The estimation measure uses collective error, while the hidden-profile task tests whether discussion surfaces unshared facts; pre-to-post estimation change barely discriminated conditions.Source: §3.4.2 · p. 11

3.5 Artifact controls

• Artifact controls address response-style priming with a neutral filler topic and address truncation by reconstructing realized trait distributions for every run.Realized trait dispersion is entered as a covariate when interpreting trait-level effects.Source: §3.5 · pp. 11–12

4.1 Models and serving

• Six open-weight instruction-tuned models spanning 3–32 billion parameters test family and scale, but the unbalanced design only partially separates them and is not a benchmark.The models were selected to run on one consumer GPU; two completed models were excluded after failing induction validation.Source: §4.1 · p. 12

4.2 Society configuration

• Societies normally contain 100 agents over 30 rounds on an initial Barabási–Albert network, with a scale ablation at 200 agents.The authors justify 100 agents as sufficient for stable network statistics while keeping runs feasible on one consumer GPU.Source: §4.2 · p. 13

4.3 Materials

• Materials include the 120-item Big Five inventory, contested and neutral topics, screened estimation items, and a hidden-profile task; human-calibrated traits use a self-selected, nonrepresentative sample.Estimation items required solo median relative error of at least 15% and log10-estimate standard deviation of at least 0.02.Source: §4.3 · pp. 13, 15

4.4 Experimental conditions

• The 991-run programme varies trait levels and heterogeneity, crosses Openness with Agreeableness, and includes scale, topic, and cross-model replications.The design also includes a pre-specified robustness audit and an internal replication with independent seeds.Source: §4.4 · pp. 15–16
• Circularity ablations remove probe anchors or feed opinion-proximity, while an alternative recommender uses agents’ most recently posted stances.The audit tests sign stability under design perturbations rather than statistical significance.Source: §4.4 · p. 16, §4.5 · p. 17

4.5 Statistical protocol

• Matched-seed paired tests, mixed-effects models, within-model standardization, and family-wise Holm–Bonferroni correction structure inference; audit contrasts are assessed for sign stability.A pre-specified extension rule raises eligible contrasts from five to eight seeds, extending all conditions in affected families.Source: §4.5 · pp. 16–17

4.6 Validity commitments and implementation

• Validity commitments include validated heterogeneous profiles, real-feed interaction, bounded memory, open-ended prompts, and a human-calibrated condition; code and run-level data are released for reproduction.Integrity safeguards block runs without sourced normative statistics and abort batches with excessive failed model calls.Source: §4.6 · pp. 17–18

5.1 Manipulation and artifact checks

• Trait induction passed its prespecified gate, and the artifact checks indicate that recommender proximity, probe anchoring, response-style priming, and truncation did not generate the reported findings.Targeted and measured traits correlated at ρ = 0.93 on average against a gate of ρ ≥0.60. Removing recommender proximity preserved 33 of 36 effect signs; removing probe anchors preserved 29 of 36, while increasing extremity, and neutral-topic variance was unrelated to contested-topic extremity (r = −0.061, p = 0.33).Source: §5.1 · pp. 18–20

5.2 Trait levels shape how societies disagree

• Agreeableness and Conscientiousness reduce cross-cutting interaction, while Agreeableness and Openness reduce opinion variance as extremity rises; Neuroticism instead raises both variance and extremity.Agreeableness lowers opinion variance by 0.476 and cross-cutting by 0.111 while raising extremity by 0.213; Neuroticism produces the largest extremity increase, 0.234. The six-topic results also overturn the earlier raised-versus-lowered trait asymmetry.Source: §5.2 · pp. 21–22, Figure 6 · p. 23

5.3 Effects hold across a wider set of topics

• Across six topics, principal trait effects strengthened, but five effects genuinely reversed; immigration responded weakly rather than in opposition to the other topics.Composition explained 53.2% of opinion-variance variation, compared with 12.3% for topic. The five reversals are treated as findings of the more reliable six-topic design, and the two-topic disagreement was driven by immigration’s low sensitivity.Source: §5.3 · pp. 24–25

5.4 Heterogeneity is the strongest compositional variable

• Personality heterogeneity increases opinion dispersion while reducing segregation; homogeneous societies instead show lower variance, less cross-cutting interaction, and greater echo-chamber closure.The homogeneous condition lowered opinion variance by 0.783, reduced cross-cutting by 0.120, and raised echo-chamber closure by 0.334. The human-calibrated condition was the most insular and uniquely showed collective-intelligence deficits, including collective accuracy −0.275.Source: §5.4 · p. 26, Table 5 · p. 26

5.5 Agreeableness moderates the effect of Openness

• Agreeableness determines Openness’s effect on opinion variance: raising Openness increases variance in disagreeable societies but decreases it in agreeable societies.The crossover is supported across replication models, whereas the primary model’s estimate is influence-driven; the cross-cutting interaction is directionally consistent but not established in the primary model.Source: §5.5 · pp. 27–28, Table 6 · p. 28

5.6 Collective intelligence: one measure works, one fails

• Post-discussion accuracy distinguishes conditions, but pre-to-post accuracy change shows no between-condition signal and hidden-profile accuracy is near a floor.The change measure had F = 0.35 and p = 0.995; private factual estimates barely changed during discussion. Hidden-profile accuracy was near 0.12 across conditions in the primary model and near zero in the replication model.Source: §5.6 · p. 28

5.7 Two further measures, reported for completeness

• Opinion assortativity on the follower network did not discriminate between conditions, with no condition shifting the mean by more than 0.024.The measure averaged −0.022 across primary-model runs and was reported as a failed measure rather than omitted.Source: §5.7 · p. 29
• Sarle’s coefficient shifts with Agreeableness and human calibration, but its associations with extremity and cross-cutting suggest it adds no distinct polarization dimension.It rises under high Agreeableness (+0.121) and the human-calibrated condition (+0.122), and falls under low Agreeableness (−0.091); it correlates with extremity at r = +0.55 and cross-cutting at r = −0.49. It is undefined when a society converges completely.Source: §5.7 · p. 29

5.8 Cross-cutting contact is associated with collective accuracy

• The adjusted cross-cutting association contrasts with three unadjusted accuracy associations that do not survive adjustment, while hidden-profile accuracy is equivalent to no meaningful association with polarization.Unadjusted correlations were cross-cutting r = +0.41, opinion variance r = +0.27, echo-chamber closure r = −0.23, and extremity r = −0.20. After adjustment, the latter three are not significant; all four hidden-profile associations are smaller than the |r| = 0.20 equivalence bound, a result qualified by a floor effect in that task.Source: §5.8 · pp. 29–31
• Cross-cutting interaction is the only polarization measure whose association with collective accuracy survives adjustment; no measure predicts poorer performance after the aggregation identity is removed.The adjusted cross-cutting association is r = +0.240, p = 0.0001; the within-condition result is suggestive but inconclusive (mean r = +0.107; 16/24 conditions positive; p = 0.18). The design cannot distinguish an effect of contact from composition-related calibration.Source: §5.8 · pp. 29–30, Figure 9 · p. 30

5.9 Robustness to design perturbation

• Every tested polarization effect was stable across perturbations, while collective-intelligence effects were not.High Openness lowered opinion variance and cross-cutting in all nine cells and raised echo-chamber closure in all nine; high Neuroticism raised opinion variance, extremity, and closure in all nine. High Openness’s extremity effect held in eight of nine cells, with the exception under a model-family change. Accuracy held its sign in six or seven of nine cells, and hidden-profile accuracy in two to five of nine.Source: §5.9 · p. 31, Table 7 · p. 32

5.10 Replication across model scale and family

• Among the four admitted models, the interaction appears in three; the primary model has b = −3.86, p = 0.030, while Qwen2.5-7B shows no interaction and additive trait effects.The 14B estimate is b = −3.25, p = 0.005, and the 32B estimate is b = −0.98, p = 0.017; relative to baseline, homogeneous and diverse societies show opinion-variance changes of −0.78 and +0.31 in the primary model, −0.41 and +0.11 in Qwen2.5-14B, and −0.07 and −0.27 in Qwen2.5-7B, respectively.Source: §5.10 · pp. 31–32
• Trait-level main effects largely do not replicate, and extremity is not a trustworthy polarization measure in the replication model.Sign agreement across ten conditions was 6/10 for opinion variance, cross-cutting, and echo-chamber closure, and 4/10 for extremity. Agreeableness on cross-cutting replicated closely (−0.111 versus −0.106), but extremity signs mostly reversed and tracked filler-topic variance in the replication model (r = +0.367, p < 10−4), unlike the primary model (r = −0.061, p = 0.33). Absolute metric levels also differed substantially between models, so magnitudes do not transfer.Source: §5.10 · pp. 34–35, Table 9 · p. 35

6.1 What the results say

• Personality heterogeneity raises opinion dispersion while lowering camp segregation, whereas high Agreeableness lowers dispersion while raising segregation.A single polarization statistic can therefore imply opposite conclusions about the same system.Source: §6.1 · p. 35
• Agreeableness determines whether raising Openness increases or decreases opinion variance, and this crossover replicated in a second model family.This interaction may explain why studies manipulating Openness in populations with different Agreeableness report opposing effects.Source: §6.1 · p. 35
• No polarization measure predicts poorer collective performance; cross-cutting interaction alone retains an association with accuracy after adjustment, but causality remains unresolved.The association varies between compositions, while the within-condition test is suggestive but not significant.Source: §6.1 · p. 36

6.2 Relation to prior work

• Unlike uniform populations that converge toward agreement in Cau et al. (2025), high-Neuroticism societies sustained two camps while high-Agreeableness societies converged.The hidden-profile task also reproduced the established preference for commonly held information over uniquely held information (Stasser and Titus, 1985).Source: §6.2 · pp. 36–37

6.3 The filler-topic control earned its place

• A neutral pizza-topping probe exposed response-style variation in the replication model, making its extremity scores unreliable rather than evidence against the underlying phenomenon.The filler topic requires one additional probe per round and distinguishes a substantive effect from a prompt-sensitivity artifact.Source: §6.3 · p. 37

6.4 Limitations

• Findings characterize simulations, not human societies; topic effects vary, and the design leaves longer horizons and other feed algorithms untested.Immigration was less affected by composition than the other five topics, and the human-calibrated condition is only one plausible human-like composition.Source: §6.4 · pp. 37–39
• Questionnaire-based induction does not establish behavioral validity, lexical-marker scoring was too circular to report trait drift, and one intelligence measure failed while another neared floor.The interaction replicated across three models, but the small, unbalanced model sample does not establish generality; the contact–accuracy association also lacks a demonstrated mechanism.Source: §6.4 · pp. 38–39

6.5 Implications and next steps

• Researchers should report and control personality composition; platform studies should jointly vary composition and feed design, whose interaction remains untested.The authors also recommend routine neutral-topic controls, which caught a model-specific artifact.Source: §6.3 · p. 37, §6.5 · p. 39

7 Conclusion

• Across 991 simulations, composition shaped polarization, but no polarization measure predicted poorer collective performance.Heterogeneity raised opinion dispersion while lowering segregation; Agreeableness moderated Openness, and cross-cutting interaction alone retained an adjusted association with accuracy.Source: §7 · p. 40
• A neutral filler topic exposed a response-style artifact, while robustness checks separated six stable polarization effects from collective-intelligence effects that did not survive perturbation.The authors encourage routine adoption of these inexpensive controls.Source: §7 · p. 40

Supplementary information.

• The supplementary file contains full condition-by-metric results, tested contrasts, robustness-audit values, item-screening records, and validity-principle compliance information.Source: Supplementary information · p. 41

Availability of data and materials.

• Run-level results for all 991 simulations and derived analysis tables are available through Zenodo and the project repository.The simulations use public input data, and the project releases its framework and analysis pipeline.Source: Availability of data and materials · p. 41

Competing interests.

• The author declares no competing interests.Source: Competing interests · p. 41

Funding.

• The work was supported by a research grant from the German University of Digital Science, whose funder had no role in the research or publication process.Source: Funding · p. 41

Authors’ contributions.

• RBT was the sole author and handled the study’s conception, methodology, implementation, experiments, data, visualizations, and manuscript.Source: Authors’ contributions · p. 41

Intelligence in LLM-Based Social Simulations

• Supporting material referenced in the main article is available in the accompanying file.Source: Intelligence in LLM-Based Social… · p. 45

• The IPIP-NEO-120 was administered to every persona configuration before any experimental run.Source: S1 Induction validation · p. 47
• Trait induction was assessed with the Spearman correlation between targeted and measured trait levels.Source: S1 Induction validation · p. 47
• The pre-specified gate required a mean correlation of at least 0.60.Source: S1 Induction validation · p. 47

• The trait matrix used in every completed run can be reconstructed exactly from its condition and random seed.Source: S2 Realised trait distributions · p. 47
• Realised dispersion narrowed at extreme trait means because of sampling-distribution truncation; the analysis controlled for this and bounded its effect at roughly one twelfth of the largest effect.Source: S2 Realised trait distributions · p. 47
• The realised means and dispersions are shown by condition, and the intended trait moved in each condition.Source: S2 Realised trait distributions · p. 47, Figure S1

S3 The society under lowered traits

• Figure S5 shows that lowered-trait societies remained mixed: none reproduced raised Agreeableness convergence or the stable split under raised Neuroticism.Across conditions, raising a trait produced ten corrected-significant contrasts, whereas lowering one produced none.Source: S3 The society under lowered traits · p. 47, Figure S5

S4 Discussion topics, verbatim

• Agents rated agreement from −3 to +3; the first two statements were used throughout, and the remaining four were added for the six-topic replication.Source: S4 Discussion topics, verbatim · p. 48
• Pineapple on pizza served as a neutral control, excluded from polarization measures and used only to detect response-style artifacts.Source: S4 Discussion topics, verbatim · pp. 48–49
• The topics included gun control, immigration, and a carbon tax.Source: S4 Discussion topics, verbatim · p. 49
• The topics also included nuclear power, universal basic income, and age verification on social media.Source: S4 Discussion topics, verbatim · p. 49

S5 Effect on opinion variance, by topic

• The topic-specific opinion-variance table reports each condition’s difference from the six-topic baseline by topic; the main article reports the average across columns.Rows show consistency across conditions, while columns show topic responsiveness to composition.Source: S5 Effect on opinion variance, by topic · p. 49

S6 Two-topic and six-topic estimates compared

• Two-topic and six-topic estimates were strongly correlated for opinion variance (+0.90), cross-cutting interaction (+0.93), extremity (+0.87), and echo-chamber closure (+0.67).Sign agreement was 82% for contrasts exceeding 0.10 in absolute value and 61% below it.Source: S6 Two-topic and six-topic estimates… · p. 49

S7 Models excluded by induction validation

• Two models were excluded before interpretation: Qwen2.5-3B failed the Agreeableness rank criterion (ρ = 0.48 vs. 0.60), and Llama-3.2-3B failed the magnitude criterion.Llama’s manipulation moved measured scores by 0.35 and 0.71 points, versus 2.22 and 3.06 in the primary model; its opinion-variance condition-mean range was 0.09 versus 0.45–2.06 in every admitted model.Source: S7 Models excluded by induction… · p. 50

S8 Induction validation, by model

• For each model and each of nine response-surface configurations, induction validation used rank correlation and the measured-score difference between the highest and lowest targeted levels.Both criteria were applied to each manipulated trait.Source: S8 Induction validation, by model · p. 51

S9 Further cross-model material

• The supplementary Openness-by-Agreeableness surfaces cover opinion variance and cross-cutting interaction in the primary and first replication models; only one model showed cell-level significance for cross-cutting interaction.The sign-agreement figure covers trait-level main effects, which the main text describes as close to chance and does not claim as replication.Source: S9 Further cross-model material · p. 51, Figure S8 · p. 52, Figure S7

S11 Counts behind the segregation rates

• Under raised Agreeableness, agents exchanged more replies while over 4,000 ordered pairs still held opposing opinions, so the low cross-cutting rate was not due to a collapsed denominator.The counts were recovered from runs at seeds outside the experimental design.Source: S11 Counts behind the segregation rates · p. 53

S12 Results ledger

• The results ledger reports 991 analysed runs, Cross-cutting vs collective accuracy r = +0.409, and Heterogeneity dose-response r = +0.966, n = 24.Collective-intelligence composites are standardised within model and measurement regime.Source: S12 Results ledger · p. 53

S13 Run accounting

• Run counts are itemized by experiment, model, topic set, and condition, and sum to the total reported in the article.The accounting includes replications and explicitly marks some smaller-model replications as excluded.Source: S13 Run accounting · p. 54

S14 Complete prompt texts

• Prompts separately elicit personality ratings, social-media actions, private opinion probes, numerical estimates, and team choices using shared and private facts.The action prompt offers posting, replying, liking, following, unfollowing, or passing.Source: S14 Complete prompt texts · pp. 54–55

S15 Collective-intelligence item screening

• Six of 14 candidate estimation items passed both screening thresholds: solo median relative error of at least 0.15 and log10-answer standard deviation of at least 0.02.The rejected items included cases where agents answered almost identically, leaving little for aggregation to do.Source: S15 Collective-intelligence item… · p. 55

S16 Complete results by condition

• The condition-level results report means over seeds for opinion variance, extremity, assortativity, echo closure, cross-cutting, collective accuracy, hidden-profile accuracy, and median error.A complete run-level file with all metrics and seeds is released with the code.Source: S16 Complete results by condition · pp. 56–57

S17 All contrasts against baseline

• In replication-model families, 33 hidden-profile contrasts had exactly zero paired differences across seeds, leaving test statistics and effect sizes undefined rather than unknown.The appendix reports these differences as 0.000 instead of imputing values.Source: S17 All contrasts against baseline · p. 60

S18 Robustness audit, per perturbation

• The robustness audit compares each anchor condition with the baseline under matching perturbations; the main text reports reference values, ranges, and sign-stability fractions.The per-perturbation table includes unperturbed and perturbed estimates.Source: S18 Robustness audit, per perturbation · pp. 81–82

S19 Full response surfaces

• Across all four polarization measures and both model families, Openness moderation reverses or steepens depending on Agreeableness.The response surfaces show cell means across the Openness-by-Agreeableness grid.Source: S19 Full response surfaces · p. 83, Figure S3

S20 The hidden-profile floor in the replication model

• The hidden-profile task discriminates weakly in the primary model and not at all in the replication model, so the alignment result relies on accuracy rather than hidden-profile solving.The figure reports hidden-profile accuracy in every run by model family; replication-model accuracy sits at an absolute floor.Source: S20 The hidden-profile floor in the… · p. 83, Figure S4 · p. 84