Behavioral psychology of LLMs: Better task guidance through punishment and reinforcement

Abstract

Recent advances in Large Language Models (LLMs) have sparked interest in their emotional intelligence, but their behavioral psychology—particularly emotion-driven decision-making akin to humans— remains poorly understood. This study introduces Behavioral Consequence Scenarios (BCSs), a structured framework simulating psychologically salient punishment and reinforcement dynamics. We test the emotional perception and empathy abilities of LLMs using psychological scales and examine the emotional variation patterns across different BCSs. The findings suggest that LLMs exhibit rudimentary emotional perception and empathy capabilities, and their emotional responses generally align with human behavior. Building on this, we propose Behavioral Consequence Scenario Prompting (BCSP), a method that injects BCS-driven emotional incentives into task instructions. Experiments on Wizard-of-Wikipedia and CMMLU demonstrate BCSP’s efficacy: in dialogue generation, BCSP achieved at least a 3.53/1.74 improvement in BLEU-2/4, and in the multi-task language understanding test, BCSP gained a 2 % improvement in overall accuracy. Our findings bridge LLMs’ alignment with human-like emotion-behavior dynamics, and encourage deeper investigation into the behavioral psychology and emotional patterns of LLMs.