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 has introduced the Recurrent Looped Transformer, a new architecture that maintains hidden states and attention caches across both prompt and response tokens without resetting. By allowing the model to retain internal data across an entire sequence, the design aims to overcome the fixed constraints of standard transformer architectures. This approach potentially enables models to achieve effectively unbounded temporal depth while processing tokens.