The Recurrent Looped Transformer (RLT) introduces a causal encoder paired with a recurrent decoder that preserves its final hidden state and layerwise sliding-window attention cache across every token. This architecture merges latent reasoning with unbounded temporal depth, enabling model-hardware and model-RL algorithm co-design. A parallel encoder handles known-token batching while the decoder extends continuous latent computation as sequences grow.

Three core principles guide the design: reasoning depth that grows with sequence length, parallel hardware work around a recurrent core, and a single transition from sampling to RL replay. The system reuses weights and memory with checkpointing, maintaining reference computation without state reset at serving boundaries. Mathematical formulations define the hidden state as H_t = (s_t, C_t^D), where the recurrent output and decoder KV carry forward through merges.

A sliding-window attention mechanism retains recent keys and values, with prompt and response sharing one complete state transition. Configurations feature 48 encoder and 48 decoder layers with shared attention and FFN weights, where each token traverses 48t decoder blocks. The design supports infinite depth as an extensible temporal path rather than infinite computation per token, though realized reasoning gains and hardware efficiency remain under investigation.