Princeton researcher introduces recurrent looped transformer to enhance token processing efficiency

A Princeton researcher, Yifan Zhang, has introduced the Recurrent Looped Transformer (RLT), a novel architecture designed to improve the efficiency of decoder-only large language models (LLMs). The RLT carries the decoder's final hidden state and its layerwise sliding-window attention (SWA) cache across tokens, eliminating the need for resets at the boundary between prompt and response. This approach allows the model to retain historical information across tokens, potentially enhancing its ability to process sequences with unbounded temporal depth. The RLT pairs a causal encoder with a recurrent decoder, where each token executes 96 logical blocks due to the combination of encoder and decoder operations. The architecture defines a design specification, including execution schedules and a reinforcement learning (RL) replay contract, though it does not report measured efficiency or scaling results. Pretraining involves full-sequence next-token prediction with full backpropagation through time, and multi-turn serving requires an exact prefix snapshot to maintain state consistency across interactions. Zhang's work builds on existing models like YOCO and DeepSeek-V4.1-Flash, while introducing unique features such as temporal feedback and depth-wise reuse. The research is detailed in a technical report, with additional resources available on GitHub and a project page.