Personality Anchoring for Social Simulation: Linking Personality, Social Behavior, and Interaction Success with LLM Agents

Abstract: Social interactions are shaped by the interplay of dispositional traits and situational context, yet systematically investigating how personality configurations between individuals jointly influence social behavior across diverse social contexts remains methodologically challenging. We address this gap by introducing a simulation pipeline adapted from the CHARISMA framework, which employs well-known movie characters and public figures as psychologically grounded agents for multi-LLM social simulation using a method we term personality anchoring. We present a large-scale empirical study examining how dyadic Agreeableness composition influences social interaction outcomes across 1,010 simulated conversations. Our results reveal a monotonic relationship between dyadic Agreeableness composition and shared goal achievement, with Homogeneous-Agreeable pairs achieving success 10 times the rate of Homogeneous-Disagreeable pairs (62% vs. 6%). Behavioral mediation analysis reveals that Agreeableness shapes goal achievement partially through cooperative strategy selection, though it continues to predict outcomes within the same dominant strategy, indicating pathways beyond observable conversational behavior. Robustness analyses confirm high consistency of results across repeated simulations (ICC = 0.89) and stable personality expression across diverse scenarios, validating personality anchoring as a viable operationalization strategy.

• The paper introduces personality anchoring to embed character traits in LLM agents, linking dyadic Agreeableness composition to joint goal success.
• It employs a CHARISMA-based pipeline to generate validated social scenarios and differentiate cooperative versus confrontational strategies.
• Empirical results reveal a strong monotonic effect, with agreeable pairs achieving 62% success compared to only 6% for disagreeable pairs.

Personality Anchoring for Social Simulation: Linking Personality, Social Behavior, and Interaction Success with LLM Agents

Introduction and Motivation

The paper introduces a robust simulation framework for systematically studying the influence of dispositional traits—operationalized via Big Five personality dimensions, focusing on Agreeableness—on social interactions between LLM-based agents. The authors address several major methodological limitations prevalent in existing computational social psychology: overreliance on trait prompting, lack of validated scenario taxonomies, insufficient exploration of dyadic personality composition, and minimal insight into the behavioral mechanisms mediating personality–outcome links. The proposed solution is “personality anchoring,” embedding personality through character-driven agent instantiation leveraging LLMs’ latent knowledge of canonical movie characters and public figures with crowd-sourced personality annotation.

Simulation Pipeline and Personality Anchoring Method

The simulation pipeline adapts the CHARISMA framework and comprises five stages: validated social goal scenario generation, curation of agent pairs based on Personality Database (PDB) Agreeableness scores, scenario expansion (easy/hard) using multi-model LLM generation, 20-turn interaction protocols with structured behavioral intent annotation, and LLM-as-a-judge evaluation of shared/personal goal achievement.

Figure 1: The CHARISMA-based pipeline systematically links human goal taxonomy, dyadic personality assignment, scenario curation, interaction, and multilevel evaluation.

Personality anchoring replaces explicit trait instructions with character identity, operationalized via PDB-based selection and controlled for inter-rater reliability. Four dyadic conditions—Homogeneous-Disagreeable (HoD), Heterogeneous-Extreme (HeE), Heterogeneous-Moderate (HeM), and Homogeneous-Agreeable (HoA)—enable granular analysis of personality composition effects. The simulation structure scales to 1,010 conversations covering 7 goal categories (e.g., Cooperation, Conflict Resolution, Competition), with careful balancing for scenario type and interaction difficulty.

Dyadic Agreeableness and Social Outcomes

The empirical analysis reveals a strong monotonic effect of dyadic Agreeableness composition on joint goal achievement: HoA pairs achieve strong shared goal success in 62% of interactions, while HoD pairs succeed only 6% of the time. The difference is stable across social goal categories, agent models, and scenario difficulty; relational scenarios (Relationship Maintenance/Building, Identity Recognition) display the largest effect sizes, while competitive goals consistently lower ceiling effects for all pair types.

Figure 2: Mean shared goal achievement increases monotonically with dyadic Agreeableness; effect sizes are largest in relational scenarios, minimal in Competition.

Cross-model evaluation confirms that these results generalize beyond a single LLM backbone, with Mistral-based agent runs nearly matching the primary results.

Behavioral Mediation and Strategy Dynamics

Detailed behavioral trace analysis uncovers that personality configuration directly shapes conversational strategy distribution. HoD pairs are characterized by high frequencies of confrontational strategies (Challenge, Dismiss), whereas HoA pairs are dominated by cooperative behaviors (Encourage, Express Gratitude, Build Consensus).

Figure 3: Top behavioral strategies by Agreeableness pair type: HoD pairs show confrontation; HoA pairs, cooperative stewardship.

Aggregation into higher-order behavioral groups clarifies the mediation mechanism: Goal achievement tracks with the dominance of cooperative behavior in dialogue. However, mediation is only partial—controlling for the dominant behavioral style, Agreeableness still predicts additional variance in outcome scores, suggesting latent pathways beyond observable intent selection (e.g., linguistic style or initiative-taking).

Figure 4: HoD pairs overwhelmingly use confrontational strategies whereas HoA pairs use cooperative/neutral strategies with higher success.

Robustness and Personality Stability

Robustness analysis via repeated identical conversation configuration runs demonstrates high ICC for outcome scores ($\text{ICC} = 0.89$), and test–retest reliability further increases with average-over-repetitions. Critically, the expression of Agreeableness remains stable for each character agent across multiple scenarios and dyadic partners, validating the personality anchoring paradigm as a psychologically reliable operationalization. Extremity clustering is observed, preserving categorical separation between low and high Agreeableness across both Conflict Resolution and Cooperation-focused scenarios.

Figure 5: Expressed Agreeableness remains stable across scenarios and partners; categorical distinctions are preserved for all agents.

Theoretical and Practical Implications

This study demonstrates that LLM-based agents, when instantiated with personality-anchored character identities, display robust and interpretable mappings from personality composition to social outcomes and strategy selection across a diverse interaction spectrum. The results empirically reaffirm social psychology’s theoretical models of Agreeableness in interpersonal conflict and cooperation but extend systematically into high-throughput, open-domain, multi-agent LLM simulation. The partial mediation by strategic behavior indicates that agent personality influences outcomes through multiple channels, some of which are not easily reducible to explicit conversational tactics.

Practically, this work offers a scalable paradigm for population-level and dyadic analysis of psychosocial variables in agent-based systems. The approach delivers granular behavioral labels and robust trait expression without dependence on brittle prompt engineering. Methodologically, the pipeline connects validated psychological instruments (e.g., human goal taxonomies, crowd-based trait annotation) with NLP agent simulation, facilitating synthetic experiments difficult to execute at human scale.

Limitations and Future Directions

The exclusive focus on Agreeableness, dependence on culturally-biased character pools, stereotyping risks in LLM character knowledge, and the limitations of LLM-as-a-judge metric constitute notable constraints. Extending the personality anchoring approach to additional trait dimensions and more demographically representative agent sets is warranted. Further work is necessary to triangulate simulated findings with human behavioral benchmarks and to probe the generalization of anchoring effects across model architectures and domains. Additionally, research on emergent phenomena in larger social groups or adaptive personality (dynamic adjustment, longitudinal drift) is a promising next step.

Conclusion

This paper offers a sophisticated, large-scale empirical examination of how LLM-embedded personality, instantiated via character anchoring, causally shapes social behavior strategy selection and dyadic goal attainment across realistic social contexts (2606.06936). The results reinforce the theoretical centrality of Agreeableness for joint success and cooperative conversational strategies but reveal that LLMs can embody stable, trait-governed personalities without explicit prompting. The presented simulation infrastructure provides a powerful foundation for future experimental social science, personality theory validation, and the design of more socially adaptive AI systems.