Building Robust AI through Causal and Social Cognition

My research focuses on developing robust and generalisable artificial intelligence systems through causal reasoning and theory of mind. Building on my recent work formalising goal misgeneralisation in causal agents and modeling theory of mind in multi-agent settings, I plan to pursue several interconnected research directions during my PhD:
Causal Exploration and Model Building: A key area I want to explore is designing agents that can efficiently build and update causal world models through strategic exploration. This builds on the framework of causal games and robust agents presented in my recent papers. Specifically, I aim to: Develop exploration strategies that allow agents to efficiently uncover the causal structure of their environment, even in settings with limited observability or interventional data, Create algorithms for rapidly updating causal models when entering new environments, with a focus on quickly identifying what interventions have occurred relative to the training distribution, Investigate how different types and quantities of data impact an agent's ability to learn accurate causal models, and how this translates to improved generalisation.
Causal Representation Learning for Robust AI: While my previous work assumed agents had access to semantically meaningful variables, in many real-world settings these variables must be identified by the agent. I plan to research how causal representation learning can be leveraged to build more robust AI systems: Explore techniques for learning causal structures and mechanisms from high-dimensional sensory data, building on recent advances in causal representation learning, Investigate the sample efficiency and generalisation capabilities of agents that learn and reason with causal models, compared to traditional deep learning approaches, Develop benchmarks and evaluation metrics for assessing the quality and usefulness of learned causal models in decision-making contexts.

Theory of Mind and Multi-Agent Reasoning: Extending my work on modeling theory of mind in multi-agent influence diagrams, I want to develop AI systems capable of reasoning about the beliefs and intentions of other agents: Design algorithms for efficiently learning and updating models of other agents' beliefs and goals through interaction and observation, Investigate how explicit modeling of higher-order beliefs impacts decision-making and coordination in multi-agent settings, Explore the connections between causal reasoning and theory of mind, and how they can be integrated to improve generalisation in social contexts.

Embodied Causal AI: To bridge the gap between theory and practice, I hope to collaborate with robotics researchers to implement causal reasoning in physical systems: Develop robot control algorithms that condition actions on learned causal world models, allowing for improved generalisation to new environments and tasks, Investigate how embodied experience and physical interaction can inform and refine causal models in ways that may be difficult in purely simulated environments, Explore how causal reasoning can enhance a robot's ability to understand and manipulate its environment, particularly in novel situations.

Towards Generalisable AI: The ultimate goal of this research agenda is to develop AI systems that can robustly generalise to new environments and tasks. By grounding learning and decision-making in causal understanding and theory of mind, I believe we can create more capable and reliable AI systems. This work has potential applications in areas such as robotics, autonomous vehicles, and human-AI collaboration.
Throughout my PhD, I plan to combine theoretical analysis, algorithm development, and empirical evaluation to advance our understanding of causal reasoning in AI.