What if a transformer never had to forget? Meet the Recurrent Looped Transformer

Princeton researcher Yifan Zhang published an architectural specification for the Recurrent Looped Transformer (RLT) on September 12, 2026: the decoder's final hidden state and sliding-window attention cache carry over to the next token without reset. It is only a specification—no measured results on efficiency or reasoning quality are included.

• Reference config: 48 encoder and 48 decoder layers with shared weights
• 96 logical blocks per token, path depth grows with sequence length
• Decoder is sequential within a sequence, speedup only via batching
• Old rollout states are not reused during training