A Transformer grew brain-like input and output gates on certain memory tasks

A Transformer trained on human working-memory tasks grew a brain-like doorman on its own. Input and output gates formed during training, deciding what got into, updated, or sent out of its scratchpad. The doorman showed up only on task distributions that benefited from the same kind of gating the human frontostriatal system uses.

That is the headline finding in a 2026 Journal of Cognitive Neuroscience paper, building on an arXiv preprint posted in February 2024. Transformer is the architecture behind ChatGPT and most modern language systems. Working memory is the brain's short-term scratchpad: the system that lets you hold a phone number while dialing, or follow the third step in a recipe after the first. The "doorman" the paper describes is a control layer that decides which items get in, which get updated, and which get sent where.

In the human brain, that control layer is largely the frontostriatal system: a circuit connecting the prefrontal cortex, which plans and decides, with the striatum, which helps select and act. Cognitive scientists have long modeled it as a gating operation, with a separate gate for input, output, and maintenance. Computational accounts attribute working memory's capacity limits to something the authors call role addressability: the ability to learn which role an item fills (the address, the value, the destination) and assign credit to the gate that handled it. In practice that means a network must, on the fly, decide that "the red square in slot 2" is one item and "move it to slot 4" is another, and that a different gate should handle each.

The question the paper asks is narrow: when a Transformer is trained on the same kind of working-memory task distributions that humans use, does it invent this gating on its own, even though nothing in its architecture tells it to? The researchers did not give the model a definition of a gate. They trained it on human-flavored working-memory tasks, then asked whether individual attention heads were behaving as if they were routing information into or out of working memory in a role-aware way.

The model's attention mechanism does develop distinct input- and output-gating strategies, but only when the training distribution actually demands them. On task distributions that do not benefit from frontostriatal-style gating, the model does not bother. The authors are careful with the wording: the constraints the Transformer learns "resemble the constraints found in frontostriatal models." They do not call it the same system. The same model can be made to grow a doorman, or not, by changing the curriculum.

That caveat is the point. The paper frames its result as evidence that role-addressable gating serves "a fundamental computational role in managing role addressability and binding," and as a bridge for future work on computational parallels between modern AI and biological models. The claim is narrower than that. The paper does not argue that Transformers and brains share an architecture, only that the same control strategy emerges in both when the training task demands it. That strategy gives the Transformer a measurable edge in storing and retrieving multiple items at once.

For the field, the practical question is whether the conditional can be tightened. If a Transformer only grows frontostriatal-like gating on task distributions that benefit from it, the design lever for AI builders is not the architecture but the curriculum. For cognitive scientists, the lever is the other way: a model that can be turned on or off with the training distribution is a more useful comparison object for human data, and the paper's authors say they want future work to dig into that. The next concrete step, per the paper, is using the same kind of training-distribution analysis to ask which other brain-like strategies the architecture can be coaxed into reproducing.