Siyuan Guo (she/her) is an AI scientist at Prior Labs and completed her PhD in machine learning jointly at the University of Cambridge and the Max Planck Institute for Intelligent Systems. Her research aims to understand intelligence and learning, with a focus on the physics of learning, causal inference, and AI for scientific discovery. In 2025 she released the Physics of Learning preprint, proposing that learning follows the principle of least action; within a week of its arXiv release the work prompted a fireside-chat invitation, multiple academic and industry talks, and collaboration discussions with infrastructure support. Her work spans theory and application, most recently advancing pre-trained foundation models for in-context causal inference in real-world scientific settings. She has published at top-tier venues such as NeurIPS and ICLR, including one oral and two spotlight presentations, and her research has been recognised by honours including MIT EECS Rising Star, the G-Research PhD Prize, the MPI-IS Outstanding Female Doctoral Student Award, and Forbes 30 Under 30 Europe.
In physics, phenomena such as light propagation and Newtonian mechanics obey the principle of least action: the true trajectory is a stationary point of the Lagrangian. In our recent work [1], we hypothesise that learning, too, follows a least-action principle. We model learning as stationary-action dynamics on information fields. Concretely, we derive classical learning algorithms as stationary points of information-field Lagrangians, recovering Bellman optimality from a reward-based Hamiltonian and Fisher-information–aware updates for estimation. This yields a unifying variational view across reinforcement learning and supervised learning, and suggests optimisers with testable properties. Conceptually, it treats the training of a learning system as the evolution of a physical system in an abstract information space.
Structure is also central to learning, enabling interventional reasoning and scientific understanding. Causality provides a framework for discovering structure from data under the hypothesis that causal mechanisms are independent. In earlier work [2], we formalise independent mechanisms as independent latent variables controlling each mechanism, and show how this perspective extends across effect estimation [3], counterfactual reasoning [4], representation learning [5], and reinforcement learning [6].
Methodologically, in collaboration with Prior Labs, we developed Do-PFN [7], a pre-trained foundation model that performs in-context causal inference — a promising out-of-the-box tool for practitioners across scientific domains.
* equal contribution · † equal supervision