When Personas Override Payoffs: Role Identity Bias in Multi-Agent LLM Decision-Making

TL;DR

This paper examines how role-based personas influence strategic reasoning in four-agent LLM systems and whether payoff visibility can overcome identity-driven behavior. Using a 2x2 factorial design across four models in an environmental policy game, the authors diagnose Nash equilibrium achievement and equilibrium selection under varying persona presence and payoff presentation. They find that role identity bias can completely suppress payoff-optimal Nash equilibria when personas are active, and that strategic reasoning only emerges when both personas are removed and explicit payoffs are provided, with strong model-dependent differences. The study demonstrates that representational choices are substantive governance decisions that shape whether multi-agent LLMs act as strategic reasoners or identity-driven actors, with important implications for deployment, disclosure, and model selection in real-world settings.

Abstract

Large language models are increasingly deployed in multi-agent systems for strategic tasks, yet how design choices such as role-based personas and payoff visibility affect reasoning remains poorly understood. We investigate whether multi-agent systems function as strategic reasoners capable of payoff optimization or as identity-driven actors that prioritize role alignment over explicit incentives. Using Nash equilibrium achievement as a diagnostic for strategic reasoning, we conduct systematic experiments across four LLM architectures (Qwen-7B, Qwen-32B, Llama-8B, Mistral-7B) in complex environmental decision-making games involving four agents. We show that role identity bias fundamentally alters strategic reasoning even when payoff-optimal equilibria exist and complete payoff information is available. Removing personas and providing explicit payoffs enables Qwen models to achieve high Nash equilibrium rates, indicating that both conditions are necessary for strategic reasoning. In contrast, personas systematically bias equilibrium selection toward socially preferred outcomes: with personas present, all of the achieved equilibria correspond to Green Transition, while models entirely fail to reach equilibrium when Tragedy of the Commons is payoff-optimal. The effect of explicit payoffs depends entirely on persona presence, revealing strong interactions between representational design choices. We also observe clear model-dependent patterns. Qwen architectures are highly sensitive to both personas and payoff visibility, whereas Llama and Mistral exhibit rigid reasoning behavior across conditions. These findings demonstrate that representational choices are substantive governance decisions that determine whether multi-agent systems act as strategic reasoners or identity-driven actors, with important implications for real-world deployment.

Figures (8)

• Figure 1: The diagram shows our multi-agent game structure with chain-of-thought reasoning. Each agent analyzes the scenario and their role identity, then selects one of two binary actions. The arrows from each agent converge to a central Strategy Profile box, representing simultaneous action selection where all four agents choose actions concurrently, and their combined choices form the strategy profile evaluated for Nash equilibrium. The dotted speech bubbles contain each agent's chain-of-thought reasoning process.
• Figure 2: Economic Scenarios Nash Equilibrium Rates by Model and Condition. This heatmap visualizes Nash equilibrium achievement rates in economic scenarios (where Tragedy of Commons is the payoff-optimal equilibrium) across all four models and four experimental conditions. The color intensity represents the Nash rate percentage, with darker blue indicating higher rates.
• Figure 3: Equilibrium Selection Patterns: Green Transition vs Tragedy of Commons (Without Personas Condition). This figure shows the distribution of Nash equilibria between Green Transition and Tragedy of Commons outcomes across all four models in the no-personas condition.
• Figure 4: Green Transition Action Selection Rates in Economic Scenarios. This grouped bar chart shows the percentage of action profiles where models select Green Transition actions (socially preferred but payoff-suboptimal) in economic scenarios where Tragedy of Commons is the payoff-optimal equilibrium. When personas are present, all models show substantial Green Transition selection rates (20-59%) despite Tragedy being payoff-optimal, demonstrating that role identity bias operates at the action level even when models fail to coordinate on Nash equilibria. Qwen models show the highest Green Transition rates with personas (40-59%), while Llama and Mistral show moderate rates (20-30%). Removing personas substantially reduces Green Transition selection for all models, with rates dropping to near-zero when payoffs are visible and personas are absent, confirming that personas drive the preference for socially aligned actions. The persistence of Green Transition selection with personas, even in conditions where Nash equilibrium rates are zero, indicates that role identity bias affects individual action choices independently of coordination success, revealing a pervasive influence of persona-driven reasoning on strategic decision-making.
• Figure 5: CoT Keyword Frequency Patterns by Model and Condition. This figure shows keyword frequency distributions (normalized per 1000 rationales) across all four experimental conditions for each model.