The Academic Salon of the Department of Psychological and Cognitive Sciences is held every Thursday afternoon. Welcome to all students and faculty members from every department!
Time: May 29 (Thursday) afternoon 14:30
Location: Room 1110, 11th floor, Lv Dalong Building
Department of Psychology and Cognitive Sciences WU Zhen's Research Group Academic Salon
Introduction
In the current era of rapid development of artificial intelligence technology, Large Language Models (LLMs) are penetrating human social life in unprecedented ways, from interpersonal communication and emotional companionship to educational counseling. While their "human-like" capabilities continue to expand, they also trigger deep thinking about their values, emotional mechanisms, and social impacts. This academic salon focuses on three major topics: prosocial behavior, gender stereotypes, and empathy perception, presenting recent research achievements of research group members. The research not only provides empirical support for understanding the prosocial behavior mechanisms of LLMs, but also provides theoretical foundation and practical insights for constructing more fair and effective human-computer interaction systems in the future.
Report One
Prosocial Behavior of Large Language Models: Value Alignment and Emotional Mechanisms
Presenter: Hao Liu
Content:
Although advanced Large Language Models (LLMs) can simulate human prosocial behavior, their degree of alignment with human prosocial values and underlying emotional mechanisms remain unclear. This study adopted the Third-Party Punishment (TPP) paradigm to compare the performance of LLM agents (GPT and DeepSeek series) with human participants (n = 100). Based on demographic and psychological characteristics, the study constructed 500 LLM agents one-to-one (100 for each model), initiated TPP games through prompt engineering, and recorded their punishment decisions and emotional responses. Results showed: (1) GPT-4o, DeepSeek-V3, and DeepSeek-R1 models demonstrated stronger fairness value alignment, with significantly higher frequency of choosing punishment options in TPP games than human baseline; (2) All LLMs reproduced the human pathway of "unfair allocation → negative emotion → punishment behavior," with DeepSeek models showing stronger negative emotion mediation effects than GPT models; (3) Only DeepSeek-R1 exhibited the human positive feedback loop of "punishment behavior → positive emotional feedback → subsequent punishment behavior"; (4) Except for GPT-3.5, other LLMs showed significant representational similarity with human high-dimensional emotion-decision patterns; (5) All LLMs exhibited rigid characteristics in emotional dynamics—compared to the contextual sensitivity of human emotions, their emotional variability was lower and inertia stronger. These findings both highlight the progress of LLMs in prosocial value alignment and reveal the importance of enhancing emotional dynamic mechanisms for constructing safe and efficient prosocial large models. These findings can not only accelerate the calibration of LLMs with human values, but also provide empirical evidence for the universality of prosocial theories in LLM agents.
Report Two
Large Language Models Amplify Empathic Gender Stereotypes: Effects on Career and Major Recommendations
Presenter: Yiqing Dai
Content:
The widespread application of Large Language Models (LLMs) in highly sensitive scenarios such as education and career counseling has raised concerns about their gender stereotype risks. However, why LLMs exhibit gender stereotypes remains unclear. This study systematically explored through three experiments whether LLMs amplify the stereotype that "women have stronger empathic abilities while men have weaker ones," and how this stereotype affects their career and major recommendations. Study 1 adopted human-machine comparison and found that six types of LLMs showed significantly higher gender bias than humans across emotional, cognitive, and motivational empathy dimensions. Study 2 manipulated input language (Chinese/English) and prompt identity (male/female), finding that English contexts and female roles more easily activated gender difference stereotypes in empathic abilities. Study 3 focused on career and major recommendation tasks, finding that large models were more inclined to recommend high-empathy-demand majors and careers to women and lower-empathy-demand majors and careers to men, reinforcing gender differences. The research revealed the generation mechanisms of LLMs' empathic gender stereotypes and their real-world impact in recommendation tasks, providing theoretical foundation and practical directions for bias identification and fairness optimization in artificial intelligence systems.
Report Three
Artificial Intelligence Superior to Strangers: Human-Computer Dialogue Reveals Competence Perception-Driven Empathy Advantage
Presenter: Jieru Yang
Content:
This study explored the impact of empathy providers (human/artificial intelligence) on adult individuals' empathy perception and its mediating mechanisms. Study 1 developed a three-dimensional empathy perception questionnaire and validated its good reliability and validity. Through online questionnaire collection, it was found that university student participants (n = 333) believed that human empathy providers could make them perceive higher levels of empathy across cognitive empathy, emotional empathy, and motivational empathy dimensions. Study 2 had university student participants (n = 102) engage in real-time conversations with strangers or GPT-4o and measured their warmth and competence perceptions of interaction partners. Experimental results showed that compared to strangers, participants perceived higher levels of cognitive and motivational empathy when interacting with GPT-4o; path analysis showed that competence perception played a complete mediating role between empathy provider type and empathy perception, while warmth perception, although positively predicting empathy perception, had no significant mediating effect. This study innovatively used real-time interaction methods, clarified the empathy advantages and limitations of humans and artificial intelligence, challenged the traditional assumption that humans have absolute empathy advantages, and provided methodological and theoretical foundations for understanding human-computer empathy and interaction applications.
