NeurosciencebioRxivPreprint
Parallel processing of orthogonal manifolds enables zero-shot composition in recurrent networks
Yuma Osako, Aineias Arango, Toshitake Asabuki
This paper shows that feedback geometry in recurrent networks determines whether learned computations can be recombined through zero-shot parallel composition.
DOI 10.64898/2026.06.14.732142