Imagine a self-driving car weaving through downtown traffic. To prevent an accident, it has to assess whether the pedestrian on the corner is about to step into the street. Or think about a stock-trading algorithm it must predict how human investors will respond to breaking news before taking action.
In both examples, machines need to go beyond simple calculations they must grasp human behavior. However, current general-purpose AI models like GPT or Llama aren’t designed for that task.
That’s where Be.FM comes in. Researchers from the University of Michigan, Stanford University, and MobLab created Be.FM—a new AI system short for Behavioral Foundation Model. It stands out as one of the first AI models specifically built to predict, simulate, and reason about human actions. The research is available on the SSRN preprint server.
Training Grounded in Behavioral Science, Not General Text
Unlike traditional AI models trained on broad text datasets, Be.FM is trained using data rooted in behavioral science including controlled experiments, surveys, and academic literature.
“We’re not using Wikipedia,” explained Yutong Xie, a Ph.D. student in information science at U-M and the study’s lead author. “We constructed a behavioral dataset with over 68,000 experimental data, around 20,000 survey participants, and thousands of scientific papers to teach the model how and why people behave as they do.”
This targeted training gives Be.FM a significant advantage over general-purpose AIs, which often miss nuanced or minority behaviors and struggle with complex social dynamics. In earlier work published in the Proceedings of the National Academy of Sciences, the team showed that mainstream AIs tend to replicate average human behavior, failing to reflect the full spectrum of human diversity. In contrast, Be.FM exhibits a variety of emerging abilities not explicitly programmed by researchers that span four major areas of application.
Predicting Real-World Human Behavior with Precision
Be.FM’s most prominent capability is its knack for predicting human behavior in real-world contexts. For instance, Xie shared an example involving a banker presenting several investment options to a group. Be.FM can forecast which choices participants are likely to favor and estimate how many will choose to cooperate or take risks. This kind of behavioral prediction can support economic modeling, product development, or public policy planning allowing organizations to simulate group dynamics before committing to expensive real-world testing.
The model also has the ability to infer psychological traits and demographic details based on behavioral patterns or background data. In practice, this means it can, for example, predict whether someone is extroverted or agreeable using their age, gender, and other demographic factors or estimate a person’s age based on their personality profile. These insights could improve user segmentation, guide tailored interventions, or inform product and service design.
Since human behavior often changes depending on context such as shifts in timing, social norms, or environmental cues—Be.FM is also adept at recognizing and reasoning about these influences.
Uncovering Contextual Drivers Behind Behavioral Shifts
Take, for example, a sudden shift in user engagement within an app between January and February. Be.FM can help pinpoint what contextual factors might be driving the change whether it’s a design overhaul, seasonal effects, or altered messaging. By examining patterns across different scenarios, the model can uncover the environmental triggers behind people’s decisions.
This positions Be.FM as a potentially powerful resource for researchers, designers, and policy analysts aiming to understand the drivers behind behavioral shifts and how best to address them.
In addition, Be.FM can structure and apply behavioral science insights to streamline research tasks. Thanks to its foundation in large language model architecture, it can assist with generating research ideas, summarizing academic literature, and tackling practical challenges in behavioral economics.
For academics and professionals alike, Be.FM could serve as a tool for hypothesis generation, study design, and even scenario modeling ahead of real-world experiments.
Outperforming Leading AI Models in Human Behavior Alignment
Be.FM consistently outperformed commercial and open-source models like GPT-4o and Llama across these four core areas, aligning more closely with human behavior. It proved especially accurate in tasks such as personality inference and behavioral scenario simulation, producing results that better reflected actual population-level trends.
However, Be.FM still has limitations it hasn’t been tested for effectiveness outside these domains. It still can’t predict major political developments or outcomes such as elections or peace negotiations.
The research team is actively working to broaden the range of domains Be.FM can cover.
“Whether it’s behavior in healthcare, education, or even geopolitics, our aim is to make Be.FM valuable in any setting where people make decisions,” said Qiaozhu Mei, a professor of information at the University of Michigan and the study’s corresponding author.
Be.FM models are available upon request, and the team encourages researchers and professionals to try them out and provide feedback.
Read the original article on: Tech Xplore
Read more: High-Security Storage Device Can Self-Destruct On Command Just Like In Spy Films