Debarghya Ghoshdastidar

For star-free languages, transformers learn best when attention weights lag behind output layer updates, defying intuitive feature learning theories.

Challenging the conventional wisdom that strong certified robustness requires heavy partitioning, this work shows how white-box knowledge of base classifiers in partition-aggregation ensembles can yield significantly tighter robustness guarantees against label-flipping attacks with fewer partitions.

Graph transformers avoid oversmoothing in deep layers by structurally preserving community information, a theoretical advantage over GCNs revealed through Gaussian process limits.