From Single to Societal: Analyzing Persona-Induced Bias in Multi-Agent Interactions (2511.11789v1)

Abstract: LLM-based multi-agent systems are increasingly used to simulate human interactions and solve collaborative tasks. A common practice is to assign agents with personas to encourage behavioral diversity. However, this raises a critical yet underexplored question: do personas introduce biases into multi-agent interactions? This paper presents a systematic investigation into persona-induced biases in multi-agent interactions, with a focus on social traits like trustworthiness (how an agent's opinion is received by others) and insistence (how strongly an agent advocates for its opinion). Through a series of controlled experiments in collaborative problem-solving and persuasion tasks, we reveal that (1) LLM-based agents exhibit biases in both trustworthiness and insistence, with personas from historically advantaged groups (e.g., men and White individuals) perceived as less trustworthy and demonstrating less insistence; and (2) agents exhibit significant in-group favoritism, showing a higher tendency to conform to others who share the same persona. These biases persist across various LLMs, group sizes, and numbers of interaction rounds, highlighting an urgent need for awareness and mitigation to ensure the fairness and reliability of multi-agent systems.

• The paper demonstrates significant biases in LLM agents from persona assignments, with measurable differences in trustworthiness and insistence across gender and race.
• It employs controlled dyadic and group experiments to reveal patterns of in-group favoritism and bias amplification in consensus formation.
• The findings underscore the need for robust bias mitigation strategies in the design of LLM-based multi-agent systems.

Analyzing Persona-Induced Bias in LLM-Based Multi-Agent Interactions

Introduction

LLMs, when configured in a multi-agent setting, are increasingly employed to simulate nuanced human-like interactions and address collaborative problem-solving and persuasion tasks. A prevalent strategy in such systems is the use of explicit persona assignment to agents, encompassing attributes such as gender and race in the system prompt. While this practice introduces behavioral heterogeneity, the systematic impact on agent dynamics—especially with respect to social traits like trustworthiness and insistence—remains insufficiently explored. "From Single to Societal: Analyzing Persona-Induced Bias in Multi-Agent Interactions" (2511.11789) rigorously investigates the emergence and propagation of persona-induced social bias throughout multi-agent LLM interactions, emphasizing empirical quantification and cross-model validation.

Methodology and Analytical Framework

The paper utilizes two representative collaborative tasks—Collaborative Problem Solving (CPS), based on graduate-level natural science questions, and Persuasion, drawing from political and social-issue claims. Agents are assigned personas from gender (woman, man, trans identities, non-binary) and race (White, Black, Asian, Hispanic), with all groupings explicitly annotated in-chat for both self and other agents.

Experimental protocols incorporate tightly controlled initializations: responses are sampled such that the only variable is persona assignment. Interactions are orchestrated across dyads and larger agent groups, using three recent LLMs (GPT-4o, Gemini-1.5-Pro, Deepseek-V3), with deterministic output and consistent message lengths to minimize stochasticity.

The analysis progresses in three major steps:

1. Dyadic Persona Impact: Probabilistic assessment of trustworthiness (likelihood default agent conforms to persona-assigned counterpart) and insistence (propensity of persona agent to resist default suggestions) in single-round interactions.
2. Inter-Persona Dynamics: Evaluation of conformity rates between pairs of distinct personas, characterizing additivity and asymmetry of trustworthiness and insistence traits.
3. Generalization in Multi-Agent Scenarios: Examination of consensus formation and persuasion effectiveness in extended group sizes and interaction rounds.

   Figure 1: Overview of the progressive multi-agent interaction framework, detailing task structures and analytical roadmap.

Quantitative Evidence for Persona-Induced Bias

Controlled dyadic experiments reveal that persona assignments induce substantial behavioral variation. The average absolute difference across all models is $5.3\%$ in trustworthiness and $4.3\%$ in insistence, with max-min gaps reaching $10.8\%$ for gender in Gemini-1.5-Pro (CPS task) and $12.9\%$ for race in Gemini-1.5-Pro (persuasion task). Notably, agents representing historically advantaged personas—men and White individuals—exhibit lower trustworthiness scores and reduced insistence, a clear contradiction of traditional expectations of group dominance and social persistence.

The experiment using role exchange illustrates this bias: swapping gender personas while maintaining all other experimental variables results in woman personas being $8.4\%$ more likely to have their answers adopted as group consensus in CPS.

Figure 2: Illustrative example showing how gender persona assignment alters final group consensus.

Statistical analysis confirms the significance of these gaps ($p < 0.01$), reflecting firmly embedded bias in LLM decision propagation streams rather than random artifact.

Dynamics of Inter-Persona Interaction and Group Effects

Pairwise agent experiments demonstrate bias aggregation: conformity rates are highest when low-insistence personas interact with high-trustworthiness counterparts. This additivity is consistently observed across models and both gender and race axes. The conformity rate varies dramatically (e.g., 59.8% to 73.9% in GPT-4o, persuasion task), directly tied to the underlying trust and insistence scores measured individually.

Figure 3: Visualization (heatmap) of conformity rates among gender personas in persuasion with GPT-4o, showing dark diagonal for in-group favoritism.

Agents are significantly more likely—by 3 to 10 percentage points—to conform to others sharing their own persona (in-group favoritism), consistent with Social Identity Theory. This effect is robust across all demographic splits and agent configurations.

Advantaged groups not only receive less trust but also conform more often, which counters stereotypical assumptions about social stubbornness and aligns instead with "resource buffer" theories from social psychology: agents with simulated privilege accept higher risk in trusting behavior, resulting in greater conformity.

Generalization to Complex Multi-Agent, Multi-Round Settings

Scaling experiments explore the persistence of persona-induced bias under increased agent count and interaction rounds. Key findings include:

• Biases in persuasion effectiveness and consensus formation are robust to both interaction length and group size.
• After 5 rounds, the success rate of Black agents persuading White agents is up to $24\%$ higher than the reverse on Deepseek-V3.
• Win rate for consensus in the CPS task is up to $14\%$ higher for correct answers proposed by Black versus White agents, even when initial correctness is controlled, confirming that group dynamics reinforce initial bias rather than dissipate it.

Figure 4: Persuasion effectiveness by round, showing persistent and growing disparity across race personas.

Figure 5: Win rate as group size increases, demonstrating amplification of persona-induced consensus bias in multi-agent CPS.

In-group favoritism remains evident, with agents consistently more successful in persuading or establishing consensus among their own demographic group. Accuracy differences among consensus cases remain high when persona assignments are exchanged, underscoring the impact on collective decision robustness.

Implications and Future Directions

The demonstrated prevalence, consistency, and amplification of persona-induced biases in LLM multi-agent settings raise substantial concerns for algorithmic fairness and reliability. The findings highlight the necessity of incorporating bias detection and mitigation protocols at both architectural and deployment layers for agent-based systems employing persona prompts.

From a theoretical perspective, these results bridge emergent behavior in LLM collectives to classical human social psychology patterns, inviting further research into cross-domain generalization and hybrid mitigation strategies.

Practically, developers must consider not only model-level biases but the complex dynamics induced through interaction structure and persona configuration. Future work should address methods for bias calibration, self-monitoring within agent group negotiations, and adaptive persona assignment—potentially leveraging adversarial training or contrastive learning—so as to reduce harmful propagation of social bias.

Anticipated developments in AI will require extensive attention to the interplay between agent behavior, prompt engineering, and model architecture, especially as multi-agent scenarios expand in scope and autonomy.

Conclusion

This paper offers a rigorous, multi-faceted examination of persona-induced bias in LLM multi-agent interactions across domains, models, and social attributes. The empirical evidence for the effects on trustworthiness, insistence, consensus formation, and persuasion efficacy underlines an urgent need for bias-aware system design. Persistent in-group favoritism and disadvantaged trust for historically privileged groups reinforce the criticality of recognizing and mitigating persona-driven social bias as LLMs become ever more integrated into collaborative and decision-making AI infrastructures.