Princeton PhD student Yifan Zhang has introduced Recurrent Looped Transformer, or RLT, a transformer architecture that adds a persistent internal state passed from one token to the next. In a standard transformer, the next token is generated mainly by reading prior context through attention and the KV cache. RLT keeps that mechanism but adds another path: the internal state produced after one token is processed is handed directly to the following token. The report describes it as giving the model a kind of relay baton for reasoning. One token finishes, passes its current state forward, and the next token continues from there.
The number of network layers each token runs through does not change, but the state chain keeps extending as the text grows longer. The paper calls this property “infinite time depth,” meaning the computation chain has no fixed upper bound in length. The report also draws a distinction between RLT and other recurrent-style architectures such as Ouro and Astra. Those approaches mainly let the same token run through the same network multiple times, while RLT focuses on continuous state transfer between different tokens.
Princeton PhD student Yifan Zhang has proposed Recurrent Looped Transformer, or RLT, a new transformer architecture that adds an internal state carried forward across tokens.
In a standard transformer, the model generates the next token mainly by reading earlier context through attention and the KV cache. RLT still looks back at prior context, but it also passes the internal state left behind after the previous token is processed directly to the next token.
The report describes this as giving the model a kind of “relay baton” for thought. After the first token finishes its computation, it passes its current state to the second. The second updates that state and passes it to the third. Each token still runs through the same number of network layers as before, but the state chain keeps extending as the text gets longer.
The paper calls this property “infinite time depth,” referring to a computation chain with no fixed upper limit on length.
The report says RLT also differs from recurrent architectures such as Ouro and Astra that have drawn attention recently. Those methods mainly make the same token run through the same network multiple times. RLT, by contrast, keeps internal state flowing across different tokens. The comparison in the report frames the first approach as checking the same answer several times before turning it in, while the second is closer to one runner handing progress directly to the next.
This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan. Disclaimer:
The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.
Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.
