Research

Our research spans the intersection of artificial intelligence, complex systems, and computational methods. We develop novel approaches to understanding and engineering intelligent multi-agent systems.

LLM Strategizing and Theory of Mind

Theory of Mind Research

how do LLMs think, strategize & plan? How do their interactions differ when they interact with humans vs with one another?"

Large language models demonstrate impressive zero-shot reasoning abilities, but struggle significantly with long-term strategic reasoning in single or multiagent settings. Recent work shows that when placed in competitive, cooperative, or mixed-motive environments, LLMs exhibit nontrivial emergent strategic behavior.

In today's world, so dominated by either human LLM engagement or multi-LLM interaction, understanding and making such long-term reasoning robust is key to society. We combine tools from computational Game Theory, Large Language Models, finetuning, and training with concepts from cognitive social science to develop in depth understanding of such systems such things

Selected Publications & Challenges:

Chemical Artificial Intelligence

Chemical AI Research

How can AI systems advance our understanding of the natural sciences, and how can we learn from automation to autonomisation? Advancing how we conduct scientific discovery, both theoretically and experimentally, can have a profound impact on our society.

Within the Big Chemistry consortium, we tackle groundbreaking scientific problems by utilizing a combination of Deep Learning methods, Foundation models, molecular and dynamical simulations, with high-throughput experimental data

Selected Publications:

Game Theory & Deep Learning

Mean Field Games Research

How can Deep learning help us understand game-theoretical scenarios for unseen and generalized games?

Understanding players' value attribution, such as Shapley value or bandit indexes, in games is fundamental; it can help us shape coalitions, influence decision-making, and plan our strategy effectively. However, calculating such indexes is computationally intensive to the point of being intractable. Understanding game-theoretical scenarios using tools from AI is a key theme of our group.

Selected Publications:

Dynamical Aspects of Learning in Neural Networks

Training Dynamics Research

What can the underlying dynamical process of training neural networks teach us about the final state of it and its learning capacity?

Understanding how neural networks learn from data and what the underlying dynamical processes is is key to fundamental aspects of our theoretical understanding of artificial neural networks. Using deep mathematical tooling from dynamical systems, combinatorial optimization, and at-scale deep learning systems, this insight is key to our group's research.

Selected Publications: