The Recurrent Looped Transformer: When a Model's Reasoning Depth Grows With the Conversation
Last Updated: 18 September 2026
A new architecture called the Recurrent Looped Transformer lets a model's reasoning path grow with the length of the conversation instead of being fixed when training ends. After a thousand tokens of back-and-forth, the latent computation has passed through the model's core block a thousand times over, so the effective depth of the model's thinking scales with the exchange. The catch, stated plainly by its own author: the reasoning gains are claimed, not yet demonstrated.
What the Recurrent Looped Transformer Actually Does
Most language models reason within a fixed structure: the signal passes through the same layer stack once per token, and extra thinking must be written out as visible text. A recurrent-depth model instead loops one core block, updating an internal state, so the computation spent on a problem becomes adjustable rather than fixed. The new Recurrent Looped Transformer extends this across a whole conversation, pairing an encoder that remembers the exchange with a decoder that carries its state forward, so effective depth grows with conversation length. The arithmetic is simple even where the architecture is not: more exchanges, more passes through the core, more accumulated computation behind each answer.
How a fixed-depth model differs
Whether a question needs one step of reasoning or twenty, a conventional model passes through its stack exactly once per token. A reasoning mode writes its intermediate steps out as visible text because that is the only space it has to think in. The reasoning depth you paid for at training time is the reasoning depth you have forever.
What looping changes
The loop is what makes the difference: one core block, run repeatedly, updating an internal state on each pass. The Recurrent Looped Transformer, a technical report published on 12 September 2026 by researcher Yifan Zhang, applies this across a whole conversation. It pairs a causal encoder, which builds a memory of everything said so far, with a recurrent decoder that carries its final hidden state and an attention cache from token to token. The report's concrete configuration uses 48 encoder layers and 48 decoder layers with weights shared across steps, and after t tokens the latent path has traversed 48t decoder blocks. Depth, in other words, becomes a function of how long the exchange has run.
The report is careful about what this means. "Infinite depth" refers to an extensible temporal path, not infinite work inside any single token, and the report states explicitly that realised reasoning gains, hardware efficiency and reinforcement-learning scaling remain to be established. That honesty is worth more than the headline: this is an architecture proposal with a defined mechanism, not a benchmark result.
The Evidence Behind Recurrent Depth
The idea did not appear from nowhere, and the wider research behind it has real results attached. According to a NeurIPS 2025 study by Geiping and colleagues, a recurrent-depth model trained from scratch on 3.5 billion parameters and 800 billion tokens improved significantly on maths and coding tasks when given more compute at inference time, without needing chain-of-thought training data. According to a 2026 arXiv study of implicit reasoning, models trained on five-step reasoning chains generalised to ten-step problems, and reached 14-step and 19-step chains at higher training iteration counts. Recurrent depth, in controlled conditions, stopped being a hard ceiling.
Test-time compute that pays off
The NeurIPS result matters as much for what the model did not need as for what it achieved: no chain-of-thought training data, no visible scratchpad, no specialised reasoning corpus. The reasoning happened in latent space, with the model adjusting how hard it worked per token, which also means the cost profile differs from reasoning modes that bill for every visible thinking token.
Reasoning past its training depth
Generalisation is the harder test, and it is where the arXiv study is most precise. A model trained with six recurrent iterations extrapolated to reasoning chains of 14 steps, and one trained with eight reached 19, in both cases well beyond the five-step chains in their training data. A dynamic iteration strategy during training matched the extrapolation of a fixed eight-iteration budget, which suggests the training recipe matters as much as the architecture itself.
What is fixed | What recurrent depth changes |
|---|---|
Depth set by the layer stack at training time | Depth extends with inference iterations or conversation length |
Extra reasoning must be verbalised as text | Reasoning happens in latent space, invisible but computational |
Compute per token is constant | Compute can scale with the difficulty of the problem |
The Catch: Overthinking and Unproven Claims
Two limitations belong in any honest coverage. The first is documented: the arXiv study identified "overthinking", where excessive recurrence degrades predictions, with gains stopping past roughly fifteen iterations for its dynamic model. The second attaches to the new report itself: its central promise is unproven, flagged as unestablished by its own author. A mechanism with the evidence still to come is a proposal, not a result you can buy on.
Overthinking is a measured limit
Overthinking inverts the usual assumption that more inference compute always helps. Past the useful range, extra passes through the recurrent block actively degraded predictions on problems unlike the training distribution, so iteration count becomes a tuning decision with a real failure mode rather than a free dial. Any deployment of this architecture inherits that decision.
The new report's numbers do not exist yet
Because the Recurrent Looped Transformer is a technical report rather than a benchmarked release, there are no benchmark tables to inspect. It defines the mechanism and leaves the evidence to future work. Readers of AI announcements see this pattern often in reverse, where benchmark numbers arrive with no explanation of the mechanism. This one is the opposite: a mechanism with the numbers still to come.
In the discovery workshops we run, owners almost never arrive asking about architectures. They arrive with a piece of multi-step work that keeps going wrong, a quote that misses items, a report that takes a day to assemble, and they want to know whether the tools have got better at that kind of sustained, multi-stage thinking. The honest answer for this development is: the direction of travel is real and measured in the research, and the newest announcement is a proposal whose payoff is still untested.
What It Means for the Tools You Actually Use
Nothing changes in your tooling this month, and that is the correct takeaway. Recurrent-depth architectures matter to a business owner for one reason: the valuable work you would hand to AI, such as tracing a disputed invoice or checking a contract clause by clause, is overwhelmingly multi-step reasoning, and architectures that remove the ceiling on reasoning depth attack exactly that. The significance is directional. The NeurIPS result shows the approach learning on maths and coding, the arXiv result shows it generalising past its training depth, and the new report extends the idea across conversations. The next piece of evidence to watch for is benchmark results on the RLT design itself. Treat any vendor claim built on this architecture before those arrive with the scepticism you would apply to any unproven mechanism.
Sources
- [1] Geiping et al., "Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach", NeurIPS 2025:
- [2] "Loop, Think, & Generalize: Implicit Reasoning in Recurrent-Depth Transformers", arXiv:2604.07822:
- [3] Yifan Zhang, "Recurrent Looped Transformer", technical report, 12 September 2026:
