A team of scientists from the University of Michigan, Stanford University, and commercial teaching experiment platform MobLab has developed a new artificial intelligence system called “Behavioral Foundation Model (Be.FM)”, Phys.org reported on July 16. This AI model is among the first dedicated AI models focusing on human behavior prediction, simulation, and reasoning, with related research results published on the Social Science Research Network preprint server.
Unlike traditional AI that relies on general text training, Be.FM is trained using exclusive datasets from behavioral science. The research team built a behavioral database containing approximately 20,000 survey respondents and thousands of research data items, enabling the model to deeply analyze the motivations behind human behavior. It has demonstrated excellent performance in four application scenarios.
The model has the ability to accurately predict real-world behaviors. Take the banking industry as an example: when needing to assess the preferences of customer groups for different investment options, Be.FM can predict choice tendencies, willingness to cooperate, and risk tolerance, providing low-cost behavioral simulation solutions for economic modeling, product testing, and policy formulation.
Be.FM can realize two-way reasoning between psychological characteristics and demographic data. It can infer personality traits through basic information such as age and gender, and also reverse-deduce demographic attributes based on personality characteristics, which is of great value for user profile construction and personalized service design.
The model is also good at capturing the impact of environmental factors on behavior. For instance, when analyzing changes in users’ behaviors between January and February, it can accurately identify the mechanism of action of situational variables such as seasonal factors and social norms, providing a scientific basis for behavioral intervention.
As a system based on the large language model architecture, Be.FM can also efficiently integrate behavioral science knowledge, assist in completing academic work such as literature reviews and research design, and become an intelligent collaborative partner for researchers.
Test data shows that in the above scenarios, Be.FM’s performance is significantly better than mainstream models such as GPT-4o, and its prediction results are more consistent with the behavioral distribution of the real-world population. The team is committed to expanding its application scenarios, aiming to empower all fields of human decision-making.
Experts in the field of behavioral science believe that Be.FM marks a new stage in the integration of artificial intelligence and social sciences. By focusing on the complexity and dynamics of human behavior, it breaks through the limitations of traditional AI in understanding human psychology and social behavior, and is expected to provide more accurate and practical tools for fields such as market research, public policy, and social management.
The team stated that in the future, they will continue to optimize the model, expand the scale and diversity of the behavioral database, and further improve its ability to adapt to complex social environments, so as to better serve various practical needs involving human decision-making.
Related topics:
Toddler Virilised by Father’s Testosterone Gel: Safety Warnings Issued
Gen Z Men Drive Trillion-Dollar Shift in Male Health Industry
Living With Low Testosterone? Weight Loss Drugs Offer New Hope
