Can LLMs improve the accuracy of behavior prediction for personnel in high-stakes scenarios by predicting emotions? — A fine-tuning study based on scarce homogeneous cultural data in high-pressure environments

Research on human behavior prediction using LLMs is increasingly prevalent. This study addresses a key question in the interdisciplinary fields of behavioral science and artificial intelligence: Can LLMs enhance the accuracy of predicting personnel behavior by predicting emotions in high-pressure scenarios? We rigorously screened homogeneous cultural data for personnel in high-stakes roles, constructing the first multidimensional dataset of “scenario-emotion-behavior” under these conditions. Following the fine-tuning of the LLM based on this dataset, we evaluated its behavior prediction accuracy. Experimental results reveal that emotion-predicting LLMs outperform baseline LLMs and behavior-predicting LLMs on multiple metrics, emphasizing the crucial roles of the Emotion-Imbued Choice Model and Behavioral Decision Theory in enhancing LLM behavior prediction capabilities. This study promotes the interdisciplinary integration of AI with cognitive and behavioral science, offering fresh insights into high-risk behavior domains and establishing a novel paradigm where LLMs improve behavior prediction accuracy through emotion prediction.