Topological Flow Matching

A team of researchers has introduced topological flow matching, a new generative modeling framework designed to preserve the structure of complex data domains such as brain graphs and traffic networks, according to a paper submitted on 14 Jun 2026 [1]. Standard flow matching, a generative modeling technique valued for its simplicity and strong empirical performance, typically treats signals on structured spaces as points in Euclidean space [1]. This approach overlooks the rich topological features of the underlying domains [1]. Flow-based generative models explicitly model a probability distribution by transforming a simple distribution into a complex one, allowing for direct computation of likelihood [5]. The new framework, detailed in a paper by Kacper Wyrwal, Ismail Ilkan Ceylan, and Alexander Tong, reinterprets flow matching as a method for solving a degenerate Schrödinger bridge problem [1][3]. The authors inject topological information by augmenting the reference process with a Laplacian-derived drift [1]. This modification captures the structure of the underlying domain while preserving the desirable properties of flow matching: a stable, simulation-free objective and deterministic sample paths [1][4]. The work was presented as a poster at ICLR 2026 [3]. The authors describe topological flow matching as a "drop-in replacement" for standard flow matching [1][4]. The framework is designed to model distributions over signals on finite graphs and simplicial complexes [4]. The researchers demonstrated the framework's effectiveness on diverse structured datasets, including brain fMRIs, ocean currents, seismic events, and traffic flows [1][4]. The paper reports gains over standard flow matching and topological Schrödinger bridge matching on these tasks [4]. In a related development within topology-aware generation, a separate group recently introduced LATO, a framework for synthesizing explicit 3D meshes using a two-stage flow matching process that first synthesizes structure voxels and then refines voxel-wise topology features [8]. That work, which also preserves topological structure, targets image-to-mesh generation rather than the signal-on-graph domains addressed by topological flow matching [8].