A Princeton Researcher Proposes Recurrent Looped Transformer (RLT) that Carries Decoder State across Every Token, Fixing 96 Blocks per Token with Unbounded Temporal Depth

Princeton researcher Yifan Zhang proposes the Recurrent Looped Transformer (RLT), a decoder architecture that carries decoder state across tokens, enabling fixed per-token computation with unbounded temporal depth. The design merges encoder outputs with prior decoder states and retains sliding-window attention caches per layer, allowing a deep recurrent path through tokens without resetting at boundaries. The report details the architecture and execution but does not provide empirical results on