In a depth-recurrent architecture, a shared block of transformer layers is applied repeatedly to the same hidden state before producing an output, so the model can spend more computation on a hard problem without adding more visible tokens or more parameters. This differs from standard chain-of-thought reasoning, where the model externalizes intermediate steps as readable text that a human or a monitor can review. Because recurrent-depth reasoning happens inside hidden vectors rather than words, it is harder for external tools to inspect what the model is actually doing between input and output. OpenAI's Astra reportedly uses recurrent depth to improve capability and efficiency, which researchers cited by The Information flagged as a monitorability concern: a technique that helps performance can simultaneously reduce how much of a model's reasoning is legible to safety monitoring built around reading chain-of-thought text. Academic work on depth-recurrent models (such as Huginn-3.5B) has found only limited evidence that this latent recurrence functions as an interpretable equivalent to written chain-of-thought.