The architecture zoo, at a glance
All 171 trained architectures: 20 initial experts and 151 found by the search. Each card shows the token mixer, with evolved blocks marked in terracotta; open a card for the full diagram, code and metrics.
Fitness combines WT2 PPL and ListOps accuracy from the search proxies (paper formula; the SpikingDeltaNet expert scores 0.500). See “About the data” on the project page.