There is a class of problem that has quietly humiliated language models for years. Not because the reasoning is hard, exactly, but because the information isn’t text. Molecular graphs. Circuit schematics. Knowledge graphs with millions of typed edges. Scene graphs extracted from video. The structure is the meaning, and flattening it into a token sequence the way you’d flatten a spreadsheet into a CSV was always a lossy, awkward hack. The models could work with it. They just couldn’t really see it.
That’s changing, and the architecture responsible is graph-native multimodality — the integration of graph neural networks directly into the transformer stack, not as a preprocessing step but as a first-class input pathway. The newest frontier models aren’t treating graphs as serialized text. They’re encoding them with learned representations that preserve topology, then attending over those representations alongside language and images. The result is a system that reasons about structure the way it reasons about meaning.
The clearest place to see this is drug discovery. Molecules are graphs: atoms as nodes, bonds as edges, with properties that emerge from the pattern of connections rather than from any single node in isolation. For years, the standard pipeline was to encode a molecule as a SMILES string — a linearized notation — feed it to a language model, and hope the model had seen enough chemistry to reconstruct the underlying geometry from the flattened representation. It worked, up to a point. But SMILES destroys information. It can’t natively represent bond angles, spatial proximity, or the way a ring system constrains the three-dimensional shape that actually determines whether a molecule binds to a protein.
Graph-native models don’t start from the string. They start from the graph itself, using message-passing layers to propagate information across the molecular structure before it ever meets an attention head. Google DeepMind’s work on AlphaFold 3 gestured at this — extending the architecture to small molecules and nucleic acids using graph-based representations of covalent geometry. But the research frontier has pushed further, toward fully hybrid systems where graph encodings, sequence representations, and 3D structure all enter the same unified attention space, letting the model ask: given this protein pocket, what topology of molecule fits?
The same structural logic applies far outside chemistry. In chip design, circuit netlists are graphs — the logical relationships between gates, the timing constraints, the power domains. EDA tools are starting to incorporate graph-native models that don’t just optimize a placement heuristically but reason about the graph’s global structure to predict timing violations before routing is even attempted. NVIDIA’s research group has published work using graph transformers for this; the acceleration over classical approaches is substantial.
What makes this architectural shift genuinely exciting is that it’s not just adding another input modality. It’s extending the domain of things a model can think about without any conversion loss. Text is one way that meaning lives in the world. Images are another. But a huge fraction of scientific and engineering knowledge is encoded as relationships — ontologies, physical networks, dependency graphs, reaction pathways, social structures. Every one of those is a graph, and every one of them has been an uncomfortable fit for the transformer’s native format.
The scaling behavior of graph-native representations is still an open question. It’s not obvious that the same laws that govern language model scaling apply when the inputs are irregular, non-Euclidean structures. But early results suggest that as graph encoders get deeper and the pretraining corpora grow richer — more molecular libraries, more circuit databases, more biomedical knowledge graphs — the models’ ability to generalize across structural domains improves rapidly.
The frontier here isn’t just “smarter chemistry AI.” It’s models that can reason fluently across every domain where knowledge is fundamentally relational. That’s most of science. And the architecture to do it is already being built.