In a new technical report, researchers proposed stretching the computation cycle between tokens using the Recurrent Looped Transformer architecture. In this scheme, the decoder becomes recurrent for each generation step, covering both the incoming prompt and the generated response. A causal encoder forms the shared KV memory. When processing each new token, the decoder combines its representation from the encoder, its own final hidden state from the previous step, and a sliding-window cache of recent activations.
The main feature of this scheme lies in compute scaling: reasoning depth grows along with sequence length, but the processing cost for each individual token remains constant. For example, with a 48-layer decoder, the computation path after t tokens already passes through 48t decoder blocks, although each specific step requires a fixed number of operations. The exact same state transition is applied across all stages: during pre-training, SFT, response generation, and replay in RL. At the same time, the state is not reset at the boundary between the user prompt and response generation, and during RL replay, the state chain is recalculated with the model's current weights, eliminating the use of stale data from past runs.
The presented concept remains exclusively an architectural proposal for now. The authors explicitly emphasize that real gains in reasoning quality, acceleration on actual hardware, and scaling effects in RL still need to be confirmed by direct measurements in practice.
For business, this means one thing: while engineers compete on theoretically lowering inference costs, you still have to pay for infrastructure today at old rates. Architectural breakthroughs on paper do not equal ready-made enterprise solutions, and you definitely should not rush to rewrite IT budgets for recurrent transformers until the technology passes combat testing in production.