Rationality Meets Irrationality: Incentivizing Dual Cognitive Loop for LLM-Based Agent in Social Simulation

Abstract

Social simulation in artificial intelligence for social science requires agents that capture both rational decision-making and human-like behavioral dynamics. However, existing large language model (LLM)-based agents often suffer from an emotional integration gap, a separation between rational and irrational processes, and memory systems lacking emotional context. To address these challenges, we propose dual cognitive loop (DCL)-Agent, a DCL framework that integrates emotion-driven (system 1) and utility-based (system 2) processes via dynamic weighting, supported by a three-tier memory architecture. Emotions modulate goal selection, plan evaluation, and memory retrieval, enabling more adaptive and socially plausible behavior. Experiments across multiple simulated scenarios demonstrate that DCL-Agent improves success rates by 5%–25% and reduces completion time by up to 30% compared with strong baselines. Ablation studies further highlight the complementary roles of emotion, memory, and rationality.