GraViti: Graph-Level Variational Autoencoders with Relaxed Permutation Invariance

In three linesGraViti is a transformer-based variational autoencoder for entire graphs, producing a true graph-level latent space. On molecular benchmarks, the model learns to decode valid samples respecting chemical constraints. The work shows that enforcing permutation invariance can be detrimental for consistent reconstruction when a reliable canonical node ordering exists.

Summary generated by Claude — human-verified