Tag
This paper introduces addressable and cardinality-preserving global memory for message-passing neural networks via cross-attention slots, addressing the finite-capacity bottleneck of virtual nodes and improving performance on multiplicity-aware tasks.