LatentPort demonstrates cross-model transfer of a language model’s live recurrent inference state to a differently-sized sibling model without the receiver rereading the source context. Translating attention KV cache alone leaves a large performance gap, but adding a Gated DeltaNet persistent-state package lowers teacher-forced next-token loss, and a small correction module lets a 9B model beat continued 4B inference while processing zero historical prefix tokens. The authors note the evidence is limited to one architecture-matched pair and short contexts.
