It is reported that OpenAI's upcoming Astra model employs a recurrent inference architecture (recurrent depth, also known as looped transformer). This architecture allows the same piece of information to pass through the same set of Transformer layers multiple times for iterative computation before output, increasing inference computational capacity without requiring longer reasoning chains.

The Ouro model released last year by ByteDance's Seed team similarly adopted this Looped Language Model design, enabling smaller models to approach the performance of larger models through increased computation. However, part of the reasoning process occurs within the model’s internal hidden states, making it impossible for humans to fully audit whether the model has violated guidelines via external chain-of-thought (CoT) traces. To address this, OpenAI has limited the extent of recurrent depth used in Astra, preserving readable reasoning chains, and plans to implement additional CoT monitoring mechanisms.