Research

What I cannot create, I do not understand. (Richard Feynman)

I work on machines that learn like we do: by actively exploring and building models of the world, an approach I call active world modeling.

A toddler putting a new object in their mouth is running an experiment: testing what it is made of and how it behaves, and revising their model of it. State-of-the-art AI mostly learns differently, by passively absorbing vast amounts of data.

Learning in nature seems to follow general principles, on which developmental psychology, cognitive science, neuroscience and physics increasingly converge. My research bet is that building on these principles can give us AI that learns autonomously, continually and open-endedly from far less data and compute, generalizes and adapts when the world changes, and is aligned, safer and more robust.

This Machine Learning Street Talk interview discusses my research. Also available on Apple Podcasts.

Active world modeling

Active world modeling agents learn how their world works, including what kind of model best describes it, while acting to gain information and achieve goals safely. Making this work raises mathematical, statistical and engineering challenges. We are working on three interlinked questions:

Where this comes from

My work has developed three interlocking strands. First, I formalized principles of biological intelligence, such as the free-energy principle, through Bayesian mechanics: a mathematical account of how systems that persist in a changing environment can be described as inferring the world around them (Da Costa et al., 2021; Friston, Da Costa et al., 2023). Second, I helped formalize active inference, the theory of agents that follows from these principles: agents perceive and learn by approximate Bayesian inference, and plan by balancing expected utility with information gain (Da Costa et al., 2020; Barp, Da Costa, França et al., 2022). Similar principles, developed independently in cognitive science, have produced agents that reach human-level sample efficiency in simple video games (Tsividis et al., 2021).

These agents, however, rely on world models whose structure is designed by hand, which limits them to simple environments. Active world modeling, my current focus, is the main missing ingredient for scaling this approach: agents that learn not only the contents of their world models but also their structure (Da Costa et al., 2025).

ATOM

I lead ATOM Research, a focused research community developing active world modeling and scaling it toward an agent foundation model: agents that learn autonomously and continually, are sample- and compute-efficient, generalize and adapt when the world changes, and are aligned, safe and robust.

Interested in joining? Please read Opportunities and get in touch 📨.

Key papers

World models and model discovery

Active inference and reinforcement learning

The free-energy principle and Bayesian mechanics

Stochastic process foundations

The full list is on Google Scholar and in my CV*.

*Updated sparingly.