Teaching language models to understand hypergraphs natively

Most work feeding relational structures into LLMs converts them to simple pairwise graphs, losing the semantics of higher-order relationships where multiple objects connect simultaneously—like a paper written by three authors or a chemical reaction with multiple reactants. Hyper-Align converts hypergraph structures directly into tokens using a hybrid template that preserves both local connection details and global patterns, then feeds these alongside text prompts into frozen LLMs. The approach outperforms existing methods on both in-domain and zero-shot hypergraph reasoning tasks, with a new benchmark (HyperAlign-Bench) for systematic evaluation.