OpenAI’s Astra to employ ‘opaque recurrence’ that could hinder chain-of-thought monitoring

The Information reported on Tuesday that Astra, OpenAI’s forthcoming model, will use a technique called “recurrent depth” – also described as “opaque recurrence” – which enables the model to operate outside the sequential reasoning that defines existing chain-of-thought (CoT) models. The method is expected to make monitoring the model’s CoT significantly more difficult, prompting immediate concern from safety experts.

What opaque recurrence entails

In conventional CoT models, the chain of thought displays the step-by-step reasoning the model follows toward a solution. Although the representation is imperfect, it still serves as a useful tool for spotting anomalous behavior or misalignment. In the case of OpenAI’s recent agent-based activities, CoT logs were a central means of understanding what occurred. Opaque recurrence, by contrast, has the model process the same query multiple times in a loop, adopting a less linear approach. The result is fewer readable traces, effectively bypassing the standard CoT recording mechanism.

Sharp reactions from the safety community

Buck Shlegeris, chief executive officer of Redwood Research, wrote that he is “very concerned” about the report. He noted that even if Astra remains monitorable to the same extent as its predecessors, an additional push of the technique could allow a massive increase in recurrence and “completely destroy” monitoring capability. Zvi Mowshowitz, a veteran in the field, argued that legislation may be required to prevent a “race to the bottom” among labs. Ryan Greenblatt, chief scientist at Redwood, warned that opaque reasoning could expand more rapidly than visible CoT reasoning, potentially reaching a state where the model thinks almost entirely in latent space.

OpenAI’s response: limited use and commitment to monitoring

According to the report, OpenAI says Astra’s use of the technique will be limited, that CoT remains expected to be readable, and the company denied any move toward “neuralese.” Jakub Pachocki, chief scientist, posted on X that OpenAI is working to preserve and exploit CoT monitoring since the earliest reasoning models and that this “is a core goal of the current research program.” The company also announced plans for extensive CoT monitoring systems as part of its future safety initiatives.

The broader picture: industry scrutiny

It is worth noting that all models perform some degree of opaque reasoning, and few researchers treat CoT logs as a direct representation of true reasoning. These caveats do not eliminate the worry that opaque recurrence could impede monitoring, especially as its use expands across different models. A follow-up piece from The Information on Wednesday morning revealed that Anthropic and Google DeepMind are already discussing the technique actively.