Are LLMs Socially Adaptive? Contrasting Belief Evolution in Large Language Models and Humans

Researchers introduce FairMindSim, a simulation benchmark and BREM framework to evaluate how well large language models align with human ethical values through social economic games. Testing 1,017 humans against ten LLMs reveals that frontier models exhibit more human-like restraint and balanced decision-making compared to mid-tier models, which show rigid, overly punitive behavior.

This research addresses a fundamental challenge in AI development: ensuring that increasingly capable language models make decisions aligned with human values and ethics. Rather than relying on static benchmarks, the researchers created FairMindSim, a dynamic simulation grounded in social psychology that observes how both humans and LLMs adapt their behavior over time in economic game scenarios. The study's scale—1,017 human participants compared against ten different models—provides robust empirical data for understanding behavioral patterns across capability levels.

The findings reveal a counterintuitive nonlinear relationship between model capability and ethical alignment. Mid-tier models demonstrate rigid, algorithmic aggression through excessive punishment in the Third Party Punishment game, while frontier models like GPT-5 and Gemini-3-Pro show greater restraint and human-like leniency. The BREM framework successfully decomposed this behavior, showing that advanced models better reconcile competing objectives by reducing inconsistencies between stated beliefs and actual decisions.

This research has significant implications for AI deployment in real-world social contexts. As LLMs increasingly mediate human interactions, understanding their alignment trajectories becomes critical for developers and policymakers. The methodology provides a replicable protocol for stress-testing AI systems' values across scenarios, enabling more rigorous evaluation before deployment. The discovery that frontier models align better with human ethical judgment suggests that scaling capabilities alone may naturally improve value alignment, though continued monitoring remains essential.

• →Frontier LLMs demonstrate greater alignment with human ethical behavior than mid-tier models, showing more restraint and leniency in social dilemmas
• →The BREM framework successfully measures and decomposes longitudinal decision dynamics, revealing how models balance competing objectives over time
• →Mid-capability models exhibit rigid, algorithmic punishment patterns that exceed human norms, suggesting a capability-alignment nonlinearity
• →FairMindSim provides a standardized psychological stress-testing protocol for evaluating AI value alignment in controlled social scenarios
• →Advanced models reduce belief-action inconsistency more effectively, indicating improved internal coherence in decision-making processes