Researchers have released the first public benchmark for testing language models on hypergraphs, data structures that link more than two things at once.
The benchmark, called TAHB, bundles 10 real-world datasets spanning e-commerce, academia, movies, and politics, each pairing hypergraph structure with raw text. It checks two things: whether the benchmark behaves like the messy real-world hypergraphs researchers already study, and whether running that text through a language model actually helps predict outcomes on the structure. The team tested two setups - using an LLM to enrich text features before a separate model predicts, and using an LLM directly as the predictor. Combining structure and LLM-enhanced text beat baselines that used either alone.
Most graph learning benchmarks stick to pairwise edges, like one account following another, even though plenty of real relationships involve groups: a paper with five co-authors, a shopping cart with a dozen items, a committee vote. Text-attributed hypergraph learning has stayed a niche research corner mainly because there was no shared dataset to test claims against, which made it hard to separate genuine progress from cherry-picked results.
A benchmark does not make hypergraph models useful on its own, but it does mean the next paper claiming an LLM improves hypergraph predictions will have to prove it against a common yardstick instead of a private dataset.