Motivation

Many of the brain’s most impressive feats of intelligence occur not in isolation but within the context of social groups or societies. Intelligence in animals—and humans—often involves collaboration, communication, and the dynamic exchange of information between individuals. Studying isolated agents may simplify experiments, but to understand the full picture of cognition, we must explore behaviors that emerge in group contexts. In recent years, multi-agent behaviors have attracted attention across several fields, including neuroscience, statistics, and cognitive science, thanks to advancements in computer vision techniques. These new tools enable researchers to measure animal behavior in unprecedented detail, tracking their movements at high spatial and temporal resolutions.

How animals interact, share information, and adapt their strategies in complex social and environmental contexts is a key aspect of these investigations. Foraging for food is one of the oldest survival behaviors, involving intricate cognitive processes to locate food, avoid danger, and often collaborate with other animals. Understanding the underpinnings of such behavior can provide crucial insights into both animal and human intelligence.

Our study, co-authored by Rafal Urbaniak and Marjorie Xie, aimed to bridge these interdisciplinary perspectives, combining cognitive neuroscience with statistical methods to model group foraging behavior. Our goal was to construct a unifying framework that allows researchers to analyze complex group behaviors across different species and environments.

The Research Journey

This paper represents the first piece of a larger research program at Basis Research Institute with the ambitious goal of reasoning about a large class of collaborative behavior, spanning a broad range of species and contexts. Achieving this will require synthesizing data and theoretical models across many fields. The scope of the project demanded a multidisciplinary approach, involving cognitive models of decision-making, statistical methods for analyzing movement, and neural mechanisms for encoding spatial and social information. We focused on a specific but rich area of study—foraging for food, an activity that involves intense survival pressures, and both individual and collective decision-making.

Given that this would be a multi-year endeavor, we began by laying down the mathematical foundation for translating between different perspectives on foraging. To capture the diversity of approaches, we developed models capable of integrating cognitive theories, neural activity, and statistical foraging patterns. We worked with both simulated and real-world data to test our framework’s flexibility and accuracy, including data we recorded from groups of birds foraging in parks. Birds are particularly well-suited for such studies because many species rely heavily on social and collective behavior to find food, making them an ideal model for studying multi-agent foraging dynamics.

Key Results

Our key results center around translating between cognitive, neural, and statistical perspectives. In foraging, animals must assess where to move based on internal preferences (e.g., how they value food) and external cues (e.g., food location or presence of other animals). From a cognitive standpoint, animals may use an internal value function to decide on the optimal action at any moment. From a neuroscience perspective, this process involves the brain mapping environments and potential rewards, while statistical models predict movement patterns from available data. For instance, when a bird forages, the brain might generate a predictive map to estimate which locations are more or less valuable. Our framework translates this into a statistical model that can then be used to predict the bird’s movement.

By deriving analytical equivalences between these perspectives, we demonstrated that different disciplines often describe the same underlying behaviors using different terminology. For example, “The animal values food” from a cognitive model corresponds to “The brain computes a predictive map of food locations” from a neuroscience perspective, and “food is a good predictor of where the animal will move” from a statistical perspective. These equivalences allowed us to apply statistical tools to infer cognitive strategies from observational data, providing a powerful means of testing predictions about animal behavior even in natural settings where highly controlled experiments may not be feasible. This was particularly exciting for us, because our group is collecting high-resolution behavioral data from groups of birds and rats in real-world environments ranging from arctic Alaska to NYC parks and subways. 

We validated our analysis approach using both simulated data and real-world datasets, including an open-source locust foraging dataset. Our approach was able to replicate a previous study, showing how locusts integrate social information—what other locusts are doing—into their decision-making processes while they forage for food. We also collected a new dataset of multi-agent foraging birds, including wintering small bird species that often form mixed-species flocks, and used the analysis framework to analyze their different proximity preferences. These birds, such as sparrows and chickadees, rely on group behavior to improve foraging success and avoid predators, making them an excellent case study for our framework.

To facilitate future research and ensure transparency, we made our code openly available in a repository (https://github.com/BasisResearch/collab-creatures), allowing other researchers to apply our methods to their own data and expand upon our findings.

Next Steps

Our next steps will extend this framework to more species and environments. This includes analyzing animals with different movement dynamics, such as fish swimming in coherent groups, and studying foraging behavior in vastly different environments, from the Arctic wilderness to the bustling New York City subway system. We're especially interested in modeling behavior and environments probabilistically, in a way that accounts for change and uncertainty. We recently collaborated with other neuroscientists to record the behaviors of NYC rats (https://hellgatenyc.com/rat-ride-along-nyc/), and we’re very interested in collaborating with people studying different types of collaborative animal behaviors. Expanding to these varied contexts will involve developing more expressive analysis packages capable of handling broader model classes. As we adapt our models to handle these new cases, the analytical tools will need to evolve in complexity and flexibility.

This research has broad implications not just for neuroscience and animal cognition but also for fields like artificial intelligence, where multi-agent decision-making is a central challenge. The ability to infer cognitive strategies from observed behavior, particularly in group contexts, is a crucial step toward designing more sophisticated AI systems. We are also excited about the opportunities to connect neuroscience and cognitive science to the social sciences, gaining insight on social interactions. We are optimistic that this type of approach can be applied to a wide range of species and behavioral contexts, ultimately leading to a deeper understanding of both the biological and computational principles that drive intelligence in complex environments. We are always looking for new collaborations and datasets to further explore the intricate dynamics of multi-agent behavior, and we invite researchers and practitioners in related fields to reach out.