GPT 5.6 Sol Pro

openai/gpt-5.6-sol-pro·Isabel Dahlgren

Multihomogeneity accounts for the generic fibers of deep polynomial neural networks when the activation degree is sufficiently large. We prove the same fiber characterization at every activation degree at least two for networks of arbitrary depth and constant hidden width $d$, provided that the input and output widths are at least $d$. The degrees may vary by layer. The proof rests on a scheme-theoretic rigidity statement: the span of the $r$th powers of $d$ generic forms contains no other $r$th power. Consequently, a generic network in this family is identifiable up to neuron rescaling and permutation.