👈 This Machine Learning Street Talk interview discusses our research,
Also available on Apple podcasts 👉
Our group is developing an Agent Foundation Model — a general-purpose learning and reasoning agent that learns to understand novel environments through active exploration, by forming explanatory world models and executing its own experiments to refine them. To this end, there are a number of mathematical and engineering challenges, including:
Core Research Themes
• Universal modelling: Developing an increasingly expressive hypothesis space, enabling the agent to model anything from physical to social interactions.
• Model discovery: Inferring plausible models based on agent-environment interaction history, online, and under suitable priors (cf. compression & core knowledge).
• Active learning: Select actions to gain information about the world, thereby testing hypotheses through active environment interaction.
Why This Matters
• Open-ended, fully adaptive, generalist AI agents.
• Extreme sample efficiency & generalisation.
• Uncertainty quantification, safety and robustness.
• AI for scientific discovery.
The successful completion of this research program may be the key to artificial general intelligence.
Keywords
• Active model discovery, active learning & model based planning.
• Variational inference, probabilistic program induction, structure learning, causal representation learning & model based reinforcement learning.
Interested in joining?
We are looking for highly motivated students and researchers who enjoy working on mathematical theory, computational methods, engineering, or all of the above.
Interested in joining this research program? ➡️ Please read this and contact me 📨.
This is a short, non-exhaustive list from our and other research groups:
Toward Universal and Interpretable World Models for Open-ended Learning Agents
AXIOM: Learning to Play Games in Minutes with Expanding Object-Centric Models
LLM-Guided Probabilistic Program Induction for POMDP Model Estimation
Bayesian Models of Conceptual Development: Learning as Building Models of the World
AlphaEvolve: A coding agent for scientific and algorithmic discovery
Assessing Adaptive World Models in Machines with Novel Games
Human-Level Reinforcement Learning through Theory-Based Modeling, Exploration, and Planning
...