Abstract
In this paper, we investigate how psychologically grounded role descriptions influence the reasoning performance of multi-agent systems based on Large Language Models (LLMs). Building on the Pool of Experts (PoE) framework, we instantiate expert agents with thirteen distinct psychological or usercentered description frameworks-including Big Five Traits, Social Cognitive Theory, and Cognitive Load Theory-and evaluate their performances across five reasoning benchmarks. Our findings reveal that while no single framework dominates globally, statistically significant differences emerge within individual tasks, with some framings better aligned to specific cognitive demands. We use weighted non-parametric statistical tests to account for sample size imbalances, and show that the choice of agent framing systematically impacts performance in tasks requiring commonsense, symbolic, and socio-emotional reasoning. These results establish psychological agent framing as a meaningful inductive bias in agentified LLM systems and support the development of task-aware personalization strategies. All models, code, and analysis scripts are publicly available at github.com/patriziobellan86/wi-iat-2025