Princeton Researcher Proposes Recurrent Looped Transformer for Latent Reasoning

Yifan Zhang, a PhD student and AI Lab Fellow at Princeton University, has proposed a new AI architecture called the Recurrent Looped Transformer (RLT). The design aims to enable latent reasoning with unbounded temporal depth by extending the recurrent path as a sequence grows.

The RLT architecture utilizes a causal encoder for global key-value memory paired with a recurrent decoder that employs local sliding-window attention (SWA). The system consists of 48 encoder layers and 48 decoder layers, featuring shared attention and feed-forward network (FFN) weights across stages. This structure allows the decoder to carry its final hidden state and layerwise SWA cache across all prompt and response tokens.

The Shift Toward Recurrence

Most current large language models (LLMs) operate with a fixed depth, meaning they apply a set number of layers to every token regardless of the problem's complexity. Recent industry research has explored "looped" or recurrent transformers as a way to allow models to dedicate more computation to difficult tasks—essentially creating a hidden reasoning process—without linearly increasing the total parameter count. RLT formalizes this approach by integrating recurrence directly into the decoder and maintaining a consistent state across the boundary between the prompt and the response.

Implications for Model Training

One of the primary technical advantages of the RLT design is the unification of state transitions. Zhang notes that the architecture ensures a single state transition for pretraining, supervised fine-tuning (SFT), sampling, and reinforcement learning (RL) replay. By doing so, the model removes structural prompt-boundary mismatches that often complicate the training of traditional transformers.

If successful, this approach could allow models to perform more complex reasoning by extending their internal "thought process" temporally without requiring a massive increase in model size. Furthermore, unifying the state transition across the training and inference pipeline could simplify the development of reasoning-heavy models.

Remaining Questions

While the theoretical framework is established, several practical hurdles remain. Zhang acknowledges that the actual reasoning gains, hardware efficiency, and the ability for the architecture to scale via reinforcement learning have yet to be fully established. Additionally, while the architecture allows for an extensible temporal path—described as "infinite depth" in terms of sequence growth—it does not provide infinite computation within a single token.