What it is
“Recurrent depth” refers to a transformer design in which a model passes information through the same block of layers multiple times before producing an output, rather than running once through a fixed stack of layers or “thinking out loud” in generated text. Instead of scaling capability by adding more parameters or by writing longer chains of reasoning tokens, a recurrent-depth model scales by looping — iterating a shared block of computation to an arbitrary depth at inference time. The technique was formalized in an academic paper, “Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach” (Geiping et al., February 2025), which built a 3.5-billion-parameter proof-of-concept model and showed that looping its recurrent block could match performance gains associated with roughly 50 billion parameters of added scale — without chain-of-thought training data or long context windows.
The practical effect is that some of the model’s “reasoning” happens inside continuous, high-dimensional latent space rather than as words. A conventional reasoning model exposes its intermediate steps as text a human (or another AI) can read — the chain of thought. A recurrent-depth model can instead process a query several times in an internal loop, arriving at an answer through a path that never gets written down as language. Commentary on this property has called it “opaque recurrence”: the reasoning is real and appears to improve results, but it leaves fewer legible traces for anyone outside the model to inspect.
As of early September 2026, the technique attracted attention because The Information reported — citing a single anonymous source — that OpenAI’s next model, internally called Astra, uses a constrained form of recurrent depth to improve cost and performance. OpenAI has not confirmed this publicly; the reporting should be read as an industry leak, not a company disclosure.
Why it matters for AI governance and narratives
Recurrent depth sits directly on a fault line in current AI safety discourse: chain-of-thought (CoT) monitoring. A widely cited cross-lab position paper, “Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety” (Korbak et al., July 2025, with more than 40 co-authors including Yoshua Bengio), argues that being able to read a model’s step-by-step reasoning is a distinctive, if fragile, safety advantage — one that architectural choices could destroy if developers aren’t careful. Recurrent depth is exactly the kind of architectural choice the paper warns about: it moves computation into a channel that current interpretability tools are not built to read.
This makes the concept a useful test case for the observatory’s meta-layer questions. The capability story (smaller models performing like larger ones, more efficient inference) and the observability story (weaker chain-of-thought legibility) are being reported as a single bundled development, not two separate trade-offs a lab consciously chose to accept. Watching which ecosystems foreground the efficiency gain versus the monitoring loss — and whether safety researchers get a public confirmation from OpenAI or continue reasoning from an anonymous leak — is itself a signal about how transparently frontier labs are willing to discuss the erosion of oversight tools they have previously endorsed.
Key facts and dates
The underlying architecture was published academically in February 2025 by Jonas Geiping and colleagues (including researchers at the University of Maryland and Lawrence Livermore National Laboratory), with code and a model released publicly. It is a general machine-learning technique, not proprietary to any one company, and has since drawn follow-on academic work (e.g., “Depth-Recurrent Attention Mixtures”) exploring variants.
The current news cycle traces to a September 1–2, 2026 report by The Information, relayed by outlets including TechCrunch and Techmeme, that OpenAI’s forthcoming Astra model uses a constrained version of recurrent depth. AI safety figures quoted in response included Redwood Research’s Buck Shlegeris and Ryan Greenblatt, and independent commentator Zvi Mowshowitz, all of whom raised concerns about the technique degrading CoT monitorability if scaled up. OpenAI chief scientist Jakub Pachocki reportedly responded that Astra’s computation depth remains close to GPT-4-era levels and that the company caps the recurrence loops specifically to preserve chain-of-thought visibility. None of these claims about Astra have been independently confirmed beyond the original single-source report.
Where to learn more
- Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach (arXiv 2502.05171) — the original academic paper defining the architecture.
- Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety (arXiv 2507.11473) — the cross-institutional safety paper explaining why CoT legibility matters and why it’s considered fragile.
- OpenAI’s new reasoning technique alarms AI safety experts (TechCrunch) — reporting on the Astra claims and safety-community reaction.
- Techmeme roundup of The Information’s original report — links to the source reporting that prompted the news cycle.