Artificial intelligence technologies face a huge challenge of understanding and predicting human behavior
, which is fundamental in many areas, such as self-driving cars that try to determine whether a pedestrian will cross the road or not, or an investment algorithm that attempts to predict the reaction of investors before making a trading decision. Currently available AI models, such as GPT and Llama, are not designed for this purpose. Therefore, researchers began to develop a model designed specifically to predict human behavior for use in areas that require the analysis of human behavior to make important decisions.
Model Be.FM to predict human behavior
FM (short for Behavioral Foundation Model) is one of the first artificial intelligence models designed specifically to predict, analyze, and simulate human behavior. Unlike models that rely on generic texts such as those available on Wikipedia, Be.FM depends on specialized behavioral data derived from scientific experiments, questionnaires, and academic studies, including more than 68 thousand participants in scientific experiments and about 20 thousand respondents to questionnaires, in addition to thousands of studies published in various journals.
“We don't provide the model with general information; we build a database that helps it understand why people behave in a certain way,” says the study's lead researcher, Yutong Xie. This approach gives Be.FM has a higher ability to read complex social cues and understand less common behaviors, known weaknesses of generic models.
Notable strengths of Be. FM
During the testing of the model in a special study, Be.FM showed emerging capabilities for which it was not directly trained, including four basic areas:
-
Predicting human behavior in real-life situations
Be.FM has notable strengths in its ability to predict the actions of humans in real-life situations. For instance, in a scenario where a banker offers several investment options to a group of people, the model can predict which option most people will choose and how many of them will cooperate or take risks. This type of behavioral prediction may support economic modeling, product testing, and public policy analysis by simulating the behavior of groups before carrying out costly field experiments. -
Conclusion of psychological features and demographic characteristics
Be.FM can infer psychological traits and demographic information based on behaviors or a set of basic data. For example, the model can estimate whether a person is socially open or tolerant based on their age, gender, and other demographic data, or even estimate a person's age based on their personality traits. This opens up the possibility of designing products that are strictly targeted to specific categories, offering effective personalized interventions. -
Understand the impact of contextual factors and surrounding conditions
Human behavior often changes in response to changes in timing, social norms, or environmental cues, and Be.FM can detect and analyze these factors. For example, if the behavior of users of an application changes between January and February, the model can determine whether this is due to a design update, a seasonal change, or a difference in the way information is displayed. By analyzing patterns in different scenarios, the model can provide insights into the environmental cues that shape decision-making, making it a valuable tool for researchers, designers, and policymakers. -
Support research and organization of behavioral knowledge
Thanks to its structure based on large linguistic models, Be.FM can summarize scientific studies, propose new research ideas, solve complex problems in behavioral economics, and even simulate research scenarios before field testing.
Limitations and prospects
The tests showed that Be.FM outperforms models such as GPT-4 and Llama in simulating human behavior, especially in predicting personality traits and simulating social situations. However, the researchers admit that the model has clear limitations; it has not yet been tested in areas such as forecasting major political events or election results.