Recently, OpenAI's latest model, GPT-6 Astra, has captured widespread attention due to its outstanding performance in 3D rendering, graphics processing, and computer operation tasks. To further bolster the model's computer operation capabilities, OpenAI has procured tens of thousands of Mac devices for reinforcement learning training. It is reported that Astra employs an innovative recurrent Transformer architecture, which effectively enhances the model's effective depth by enabling intermediate representations to traverse the same set of Transformer blocks multiple times.

There are several variants of the recurrent Transformer design. For instance, Nanbeige features a fixed number of recurrent cycles, while Universal Transformer and Ouro support adaptive stopping. Another variant, Mixture-of-Recursions, determines the number of recurrent cycles through routing. These architectural innovations can significantly elevate the model's overall performance within a constrained computational budget.

In light of concerns that recurrent Transformers might obscure reasoning trajectories, some experts contend that shorter reasoning trajectories actually signify stronger model capabilities and lower error rates, and are not inherently linked to the recurrent architecture itself. OpenAI's Chief Scientist also addressed this issue, emphasizing that the architectural modifications are not the root cause of the reduced transparency in reasoning trajectories. Furthermore, the research also encompasses related fields such as latent reasoning, the differentiation between knowledge retrieval and reasoning, recurrence under computational alignment, and full-bandwidth Transformers.