Yifan Zhang published the Recurrent Looped Transformer (RLT) architecture. It carries latent computation across every prompt and response token. A causal encoder builds global key–value memory. A recurrent decoder combines that memory with sliding-window attention and feedback from its previous final hidden state. The design targets infinite temporal depth without increasing per-token computation. Model–hardware co-design enables efficient training and inference. Model–RL algorithm co-design supports consistent state transitions across pretraining, SFT, and RL. Preliminary synthetic experiments show RLT maintains accuracy at training length but declines on longer sequences. Results are proof-of-concept and not validated at scale.

• 48 encoder layers + 48 decoder layers
• 79K parameters
• 32 operations training length
• 128 operations evaluation length