Visualize an autonomous vehicle maneuvering through urban traffic. To prevent an accident, it must assess whether the pedestrian at the intersection is preparing to step onto the road. Alternatively, think of an investment algorithm trading shares—it must predict how human traders will respond to information before executing a transaction.
In both scenarios, machines need to accomplish more than mere calculations—they need to interpret human actions. Yet, the current general-purpose AI frameworks, such as GPT or LIama, are not designed for this purpose.
Introducing Be.FM, short for Behavioral Foundation Model, a novel AI framework created by scholars at the University of Michigan, Stanford University, and MobLab. Be.FM is among the first AI systems specifically engineered to foresee, emulate, and analyze human behavior.
In contrast to conventional models that depend on generic text datasets, Be.FM is trained on data from behavioral sciences—spanning controlled experiments, surveys, and scientific research.
“We aren’t utilizing Wikipedia,” stated Yutong Xie, a PhD candidate in information science at U-M and the primary author of the study. “We constructed a behavioral dataset—including over 68,000 individuals from experimental data, roughly 20,000 survey participants, and thousands of scientific articles—to enable the model to reason about why individuals behave as they do.”
This targeted training equips Be.FM with an advantage over general-purpose AIs, which frequently overlook atypical behaviors or misinterpret intricate social signals. For example, the team’s previous research, published in the Proceedings of the National Academy of Sciences, demonstrates that off-the-shelf AIs typically emulate average human behavior but fail to capture the richness of human diversity. More significantly, Be.FM showcases a variety of emergent abilities—skills that researchers did not intentionally program—across four principal application domains.
The foremost and most apparent strength of Be.FM lies in its capacity to forecast human actions in authentic contexts. For instance, Xie illustrated a case where a banker presents several investment options to a group. Be.FM can be employed to predict which selections individuals are likely to favor and how many will choose to collaborate or take risks. This behavioral forecasting could enhance economic modeling, product experimentation, or public policy evaluation, allowing for the simulation of group behavior prior to initiating costly real-world trials.
Be.FM can also infer psychological characteristics and demographic details from behavior or background information. In practical use, this might entail deducing whether an individual is extroverted or agreeable based on their age and gender, alongside other demographic data, or estimating someone’s age based on their personality traits. This functionality could assist researchers in segmenting users more effectively, directing personalized interventions, or informing product development.
Human behavior frequently alters in response to context, like variations in timing, social norms, or environmental signals. Be.FM can aid in identifying and interpreting these influences.
For example, when user actions in an application transition from January to February, Be.FM can help pinpoint which contextual elements may be driving the change—such as a design alteration, a seasonal trend, or shifts in how information is presented. By examining patterns across different scenarios, the model can reveal insights regarding the environmental cues that influence decision-making.
This positions it as a potentially invaluable resource for researchers, designers, and policy analysts aiming to comprehend why behaviors fluctuate and how to respond effectively.
Lastly, Be.FM can organize and apply behavioral science insights to bolster research workflows. Built on a large language model framework, it can generate new research concepts, summarize literature, or tackle practical behavioral economics challenges.
For academics and practitioners, it could serve as a tool for generating hypotheses, planning investigations, or even simulating situations prior to field assessments.
In these four domains, Be.FM consistently outperformed commercial and open-source models like GPT-4o and LIama in aligning with human behavior, particularly in tasks such as personality forecasting and scenario simulation. Its predictions more accurately mirrored real-world trends, especially at the population level.
However, the model has limitations—its effectiveness beyond these four areas remains unexplored. It is not yet crafted to predict large-scale political occurrences or outcomes like elections or peace negotiations.
The research team is actively working on broadening Be.FM’s application scope.
“Behavior in health, education, and even geopolitics—the objective is to make Be.FM applicable wherever individuals make choices,” said Qiaozhu Mei, U-M professor of information and the corresponding author of the research.
The Be.FM models are accessible upon request. The team encourages researchers and practitioners to utilize the model and provide their feedback.