New Perspectives in Advancing Graph Machine Learning

Professor of Computer Science at Stanford University.

Relational Foundation Models and the End of Task-Specific GNNs

The AI stack is currently unbalanced: we have mastered in-context learning for sequences (text) and grids (images), yet the 'dark matter' of enterprise intelligence—dynamic, heterogeneous structured data—remains stuck in the era of brittle, supervised pipelines. In this talk, I will argue that the future of Graph Machine Learning lies not in better architectures for static benchmarks, but in Relational Foundation Models (RFMs) that serve as universal structural reasoners. I will demonstrate that by abandoning fixed schemas in favor of schema-invariant message passing, RFMs achieve what was previously thought impossible: zero-shot generalization across disjoint graph topologies. We will explore how RFMs utilize structural in-context learning to treat predictive tasks not as model training problems, but as subgraph pattern matching queries within a latent relational space. By decoupling representation learning from specific edge types, RFMs do for structured data what LLMs did for language—transforming the graph from a static asset into a queryable, generative engine. The question is no longer if graph foundation models are integral to the AI stack, but how quickly we can transition from crafting individual GNNs to prompting universal relational substrates.

Professor in Halιcιoğlu Data Science Institute at University of California, San Diego.

Learning generalizable algorithmic procedures via GNNs

A central challenge in modern machine learning is learning generalizable procedures that remain effective on unseen, potentially out-of-distribution (OOD) data. Such generalization depends on a complex interplay among model architectures, task structures, data assumptions, and training methodologies. In this talk, I will focus on the interaction between model architecture and task structure in the context of graph learning. We are particularly interested in two questions: Do different graph neural networks learn fundamentally different algorithmic procedures? And can OOD generalization be achieved with only finite samples? To explore these questions, I will present our initial studies using two concrete settings, graph partitioning/clustering and graph shortest-path computation, as testbeds for understanding how graph models internalize and apply algorithmic structure. This talk is based on joint work with M. Black, S. Chen, S. Dasgupta, R. Nerem.

Senior Research Scientist at Google Research.

Assistant Professor of Electrical and Computer Engineering at UC Santa Barbara.

Topological Deep Learning: Unlocking the Higher-Order Structure of Relational Systems

The world is full of complex systems characterized by relations between their components: from social interactions between individuals to electrostatic interactions between atoms. Traditional graph methods capture only pairwise interactions, such as a covalent bond between two atoms. Topological Deep Learning (TDL) goes further by modeling higher-order relations, such as a benzene ring formed by six carbon atoms. TDL enables us to extract knowledge from these richer structures: for instance, assessing whether a protein is a promising drug target. This talk will introduce the core principles of TDL, provide a comprehensive review of its rapidly growing literature, describe open-source and accessible implementations, and raise open theoretical questions for the field. All in all, we will showcase how TDL can effectively capture and reason about the real-world complex systems, while highlighting outstanding challenges and opportunities.

Professor (W3) at the University of Stuttgart and a faculty member of the International Max Planck Research School for Intelligent Systems.

Learning (Approximately) Equivariant GNNs for Simulations of Physical Systems

Physical systems can be simulated through first-principles and numerical methods, but these approaches often become computationally prohibitive at scale. Machine-learning models can complement such methods by providing scalable, data-driven surrogates that incorporate physical structure. Equivariant graph neural networks are particularly effective in this setting because they encode rotational and translational symmetries directly into their architectures, and they have become standard components in probabilistic generative models of atomistic systems. Nevertheless, training these generative models, especially when learning invariant distributions with denoising diffusion or flow matching, can be challenging due to the high variance of standard gradient estimators and the rigidity imposed by strict equivariance constraints. We present a set of methods that address these challenges. First, we introduce a Rao-Blackwellised gradient estimator that interprets symmetry-based data augmentation as a Monte-Carlo approximation of the true gradient and replaces multiple stochastic estimates with a single symmetrized estimator. This yields provably lower-variance gradients and enables more stable and efficient optimization of invariant generative models. The same framework also improves the use of data augmentation typically employed when learning invariant distributions, enhancing the statistical efficiency of diffusion-based training. Finally, we provide a formulation for learning approximately equivariant functions by treating equivariance as a constraint that can be relaxed and adapted in a data-driven manner, allowing models to retain symmetry-aware inductive biases while accommodating the variability present in realistic physical systems.