DeltaTTT: Layerwise Optimization for Nonlinear Recurrent Memory.

arXiv:2610.08553v1 Announce Type: cross Abstract: Sequential test-time training adapts a memory network through successive updates, each computing an inner-loop gradient based on the network's previous state. Intuitively, this state dependence should allow eac