eh2022-two-model-disentangled-superposed

Summary

This is a controlled experimental setup for studying how neural networks pack more features than they have neurons: the input and output sides are given a clean, one-feature-per-slot representation, while the hidden layer is forced to represent those same features in a smaller, overlapping space. The "so what" is that it gives researchers a way to isolate and measure exactly how superposition works in the middle of a network, rather than trying to observe it in a fully uncontrolled model.