Graph is a Substrate Across Data Modalities

Graphs provide a natural representation of relational structure that arises across diverse domains. Despite this ubiquity, graph structure is typically learned in a modality- and task-isolated manner, where graph representations are constructed within individual task contexts and discarded thereafter. As a result, structural regularities across modalities and tasks are repeatedly reconstructed rather than accumulated at the level of intermediate graph representations. This motivates a representation-learning question: how should graph structure be organized so that it can persist and accumulate across heterogeneous modalities and tasks? We adopt a representation-centric perspective in which graph structure is treated as a structural substrate that persists across learning contexts. To instantiate this perspective, we propose G-Substrate, a graph substrate framework that organizes learning around shared graph structures. G-Substrate comprises two complementary mechanisms: a unified structural schema that ensures compatibility among graph representations across heterogeneous modalities and tasks, and an interleaved role-based training strategy that exposes the same graph structure to multiple functional roles during learning. Experiments across multiple domains, modalities, and tasks show that G-Substrate outperforms task-isolated and naive multi-task learning methods. The codebase, model, and datasets are available at .

Many problems across science and AI involve understanding how things relate to each other: how atoms connect in a molecule, how objects interact in a photograph, or how events unfold in a news story. All of these can be captured as graphs, that is, dots (entities) linked by lines (relationships). Today's AI systems learn these graphs separately for each task, so insights from molecules cannot help with images, and the same patterns are rediscovered rather than accumulated. We ask whether relational structure can instead serve as a shared foundation across very different problems. Our framework, G-Substrate, treats graphs as a reusable scaffold rather than a disposable byproduct. It forces every task to describe its graph in a common format, and trains the model to use each graph in multiple roles. Sometimes the model builds a graph from raw input, sometimes it reasons over a given graph to answer a question. Across chemistry, vision, language, and algorithmic tasks, G-Substrate consistently outperforms methods that learn each task in isolation. This suggests that relational structure, when treated as a shared resource, can accumulate across problems instead of being thrown away. The findings point toward AI systems that grow more capable as they see more kinds of structured data.