A fix for graph neural networks' favorite shortcut just got a lot more precise.
Graph neural networks often add a single "virtual node" that every other node can reach in two hops, letting information skip long, winding paths. The problem: when every node dumps information into the same shared node, that shortcut turns into a bottleneck, a failure mode called oversquashing. Researchers addressed this by building global memory with two specific properties: addressability, so the network can pick out one of many memory slots using a compact code instead of blending everything together, and support-awareness, so a query always has a private reference point to compare against instead of reading a confused average. They implemented this with cross-attention memory slots and a constrained bilinear memory, then tested both against standard pooled virtual nodes.
The gap is the real story here. On controlled graph reasoning tasks, the addressable designs solved problems up to depth five, while pooled virtual nodes of similar or larger size only reached about 10.6 percent accuracy. That is not a marginal improvement; it is the difference between a shortcut that works and one that quietly fails as graphs get bigger, which matters for anyone using GNNs on large molecules, social networks, or recommendation graphs.
Still, these are controlled synthetic benchmarks, not messy real-world graphs, so treat the numbers as a proof of concept rather than a verdict.