alphaXiv describes a transformer that deepens computation without adding layers
In its summary of “Recurrent Looped Transformer,” alphaXiv says the decoder carries state across every prompt and response token, while a causal encoder provides reusable global memory.
TLDR
alphaXiv describes “Recurrent Looped Transformer” as pairing a recurrent decoder with a causal encoder that provides reusable global key-value (KV) memory. It says longer sequences create deeper latent computation paths without adding physical layers. The same state-transition rule is maintained across pretraining, inference and reinforcement-learning replay, according to alphaXiv.
Combined views
76K
2 Sources, first seen ago