Universal Test-Time Training

Test-Time Training (TTT) compresses context into fast weights that are updated online and queried as recurrent memory. In existing designs these fast weights are \emph{private}: every layer owns a state no other layer touches. We argue this is an arbitrary restriction and study its alternative, \emph{universal fast weights}, in which the whole depth stack shares one routed pool of states. Sharing the pool across depth makes its full capacity addressable at every forward, so the effective state a layer can reach grows with depth, capacity is allocated by routing rather than fixed at design time, and the per-layer updates act on one bank and group into a single large kernel rather than running layer by layer. We instantiate this as \emph{uTTT-Dense}, a shared dense operator; \emph{TTT-MoE}, a routed pool that stays layer-local; and \emph{uTTT-MoE}, a globally shared pool. We further observe that a universal TTT model is secretly a fully recurrent model: one state persisting across depth makes the layer index a second recurrence axis, so the $L$-layer stack unrolls into a single recurrence over (chunk, layer) pairs and state written deep at one chunk is read shallow at the next; layer-private TTT is the fixed-partition special case. Across large-chunk language modeling and LVSM-style novel view synthesis, moving fast weights from layer-private, to layer-local routed, to a globally shared pool improves long-context retrieval monotonically --- RULER accuracy from $8.5$ to $15.5$ at 124M and $23.2$ to $27.8$ at 760M, and object-level rendering from $24.6$ to $25.5$~dB --- though the advantage concentrates at shorter contexts and does not yet hold at the longest evaluated length.