AI glossary · Basic · Language models · Safety & alignment
Recurrent Depth
Recurrent depth is an approach in which an AI model passes through the same neural network layers multiple times, letting it spend more computation on hard problems without making the model much larger. It is also called a 'looped transformer'.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Models & generation
- Added
- 2026-10-25
What it is
A standard transformer model produces an answer by passing through a fixed number of layers once each, in order. The recurrent depth approach processes a problem by reusing the same block of layers several times. Because it reruns the same layers instead of stacking new ones, it can do more computation without greatly increasing the number of parameters (the numbers the model has learned). The name comes from increasing 'depth' through the number of repetitions rather than the number of layers.
How it works
When an input sentence enters the model, it is first turned into an internal state (Hidden Representation). Each time this state passes through the shared block of layers, it becomes a slightly revised state, and the result is fed back into the same block. Only after going around this loop several times does the model produce the answer as text in the final step. An appeal of this approach is that computation can be matched to problem difficulty: fewer loops for easy problems, more for hard ones.
How it has developed
In research, there is a trend toward treating recurrent depth as a separately adjustable knob in model design. The 'DeepThink' paper posted to arXiv in October 2026 proposed a framework that separately designs a store that holds knowledge, a mechanism that selects knowledge suited to the current state, and a shared 'thinking' operation that repeatedly revises the state. What this paper newly did was separate 'how much knowledge to store', 'how broadly to use it at once' and 'how many times to repeat (recurrent depth)' into distinct design variables, and lay out a path for converting an already trained general model into this kind of recurrent structure. The research is still at the stage of presenting a conceptual model.
Why it is getting attention
Industry interest grew after reports that OpenAI is applying this technique to its next-generation models. On performance, the advantage is that more computation can be spent on hard problems without enlarging the model. Safety researchers, however, are concerned about monitoring. Existing reasoning models leave the intermediate steps of solving a problem as a natural-language CoT (Chain-of-Thought), so people can look for clues of dangerous behavior, whereas with recurrent depth much of the internal computation may not be expressed in sentences.
Existing reasoning models vs. recurrent depth
Existing reasoning modelsLeave CoT in writing
- Show intermediate steps in natural language
- Easier for people to spot warning signs
Recurrent depthRepeat the same layers
- Spend more computation without greatly enlarging the model
- Internal computation may not appear as sentences
- Monitoring may become harder
공통 Both aim to spend more computation on hard problems
Things to know
Recurrent depth cuts both ways: it is a technique that improves performance and also one that can make it harder for people to inspect an AI's thought process. This has led to proposals that AI companies avoid designs that prevent reasoning monitoring. How much this technique has been used in a particular company's models is known mostly from media reports, so it should be read as separate from officially confirmed facts. It is also worth remembering that increasing the number of repetitions increases computation time and cost as well.
Example
Among the news covered on this site is a WSJ report that the firing of three OpenAI safety researchers stemmed from an investigation into leaks about 'recurrent depth'. According to the report, the investigation was triggered by a media report on a new reasoning technique in OpenAI's models, which said the technique could improve performance while making it harder for people to understand the thought process. The fired researchers denied leaking confidential information, and before their firing they had argued for preserving structures that let people monitor AI reasoning. The episode shows that recurrent depth has become a point of debate over AI monitorability, beyond being simply a performance technique.