PolyGLOT: Graph-Token Language Modeling for Unified Polymer Property Prediction and Programmable Generation

Polymer design benefits from representations that combine efficient sequence modeling with explicit access to molecular graph structure. String representations provide compact sequences but most remain sensitive to syntax and connectivity errors, whereas sequential graph generation becomes more costly as molecular graphs grow. We introduce \textbf{PolyGLOT (Polymer Graph Language of Tokens)}, which represents a linear homopolymer repeat unit as a reconstructable sequence of mostly small molecular subgraphs. A graph encoder embeds each graph token, and a common Transformer backbone is pretrained by masked-token prediction and subsequently fine-tuned for property prediction or causal generation. PolyGLOT achieves competitive property-prediction performance while learning a hierarchy from local graph chemistry to context-dependent token states and property-organized polymer representations. Its attachment-constrained decoder produces valid repeat-unit graphs by construction, and the typically small token size makes local graph operations tractable. Beyond direct generation, the same graph-token interface supports expansion beyond the fixed vocabulary, property-guided local editing, required-substructure control, and small-molecule structural-precedent analysis. These results show that PolyGLOT provides a unified and extensible representation for polymer property prediction, generation, and downstream design.