What is recurrent depth?

Definition

Recurrent depth is a model design in which information passes through selected computational layers repeatedly. Instead of moving through each layer only once, the model can loop over part of its network to perform additional internal processing.

The approach can give difficult problems more computation without requiring the model to express every intermediate step as extra text. A looped transformer is another name used for this type of recurrent computation inside transformer layers.

ELI5

Imagine reading a tricky sentence more than once before answering a question about it. Recurrent depth gives an AI model a similar chance to revisit its internal work.

For example, an easy arithmetic question might need one pass, while a harder puzzle might benefit from several passes through the same useful layers before the model answers.

Frequently asked questions

How is recurrent depth different from simply generating more tokens?

It spends additional computation inside the model before the next token is emitted, rather than writing more intermediate text into the conversation.

Is recurrent depth the same as a looped transformer?

The terms are often used for the same general idea: selected transformer layers are reused in a computational loop.

Videos explaining recurrent depth