alphaXiv(@askalphaxiv):“Recurrent Looped Transformer” This paper makes the decoder recurrent across every prompt and response token, while a causal encoder provides reusable global KV memory. Longer sequences then create deeper latent computation paths without adding more physical layers, while keeping the same state transition across pretraining, inference, and RL replay.

2026.09.13 09:23

“Recurrent Looped Transformer” This paper makes the decoder recurrent across every prompt and response token, while a causal encoder provides reusable global KV memory. Longer sequences then create deeper latent computation paths without adding more physical layers, while keeping the same state transition across pretraining, inference, and RL replay.