A new study suggests that the brain may build complex behavior in much the same way engineers build sophisticated machines: by combining reusable components rather than creating an entirely new system for every challenge. In experiments with mice and artificial neural networks, researchers found evidence that groups of neurons in the prefrontal and parietal cortices can be organized into specialized computational modules. These modules appear to perform particular operations, such as processing sensory information or maintaining memories, and can then be reused when the brain encounters new stimuli or must hold different kinds of information in mind.
The findings, reported in Nature Neuroscience, address a central question in neuroscience and artificial intelligence: does the brain rely on a modular architecture that can flexibly recombine existing computations? Artificial neural networks have demonstrated that complex tasks can be solved efficiently when specialized modules are assembled according to the demands of a particular problem. Yet it has remained unclear whether biological neural circuits use a similar strategy. The new work indicates that at least some cortical networks may operate as reusable libraries of computations, allowing animals to adapt without rebuilding their cognitive machinery from scratch.
The researchers studied mice performing a delayed match-to-sample task with delayed report, a behavioral paradigm designed to separate the stages of perception, memory and decision-making. In this type of task, an animal is presented with a sample stimulus and must retain information about it across a delay before making a response. The delay is crucial because it requires the brain to preserve task-relevant information after the original stimulus has disappeared. By observing neural activity during different stages of the task, the researchers could examine whether the same neuronal representations were dedicated to a single event or could be recruited for multiple cognitive operations.
Their analysis identified neuronal subspaces associated with two key functions. One subspace was specialized for stimulus processing, representing information about incoming sensory inputs. Another was associated with memory maintenance, supporting the retention of information during the delay period. In computational neuroscience, a neuronal subspace is a coordinated pattern of activity across many neurons. Rather than interpreting each neuron as an independent information channel, researchers can view the population as generating activity within a multidimensional space. Different directions or patterns within that space can encode distinct variables, such as stimulus identity, timing or remembered content.
The striking result was that these subspaces were not used only for the specific information linked to their original discovery. The stimulus-processing subspace was reused to represent new stimulus inputs, while the memory-maintenance subspace was reused for different types of memories. This suggests that the brain’s organization may be defined less by individual, permanently assigned representations and more by flexible computational roles. A circuit capable of processing one kind of sensory information may be repurposed for another, provided the underlying computation remains similar. Likewise, a circuit that maintains one form of memory may support another without requiring a separate dedicated network.
The researchers also found that each functional subspace was supported by a distinct cluster of neurons. These clusters were identified in both the prefrontal cortex, a region strongly associated with planning, working memory and cognitive control, and the parietal cortex, which contributes to sensory integration, attention and decision-making. The separation between clusters provides important anatomical support for the idea of modularity. It suggests that the observed computational roles were not simply abstract statistical patterns spread uniformly across the brain, but were linked to partially distinct groups of neurons that could potentially be recruited as functional units.
To test whether the proposed modules actually played causal roles, the researchers turned to artificial recurrent neural networks constrained by neural data. Recurrent networks are particularly useful for modeling cognitive tasks because their units can maintain activity over time, allowing the system to represent information during delays. By shaping these artificial networks to reproduce patterns observed in the mouse brain, the researchers created models that connected biological activity with computational mechanisms. They then simulated the silencing of specific neuronal clusters. Disrupting one cluster interfered with a particular computation, while leaving other operations less affected, a result consistent with the presence of specialized but reusable modules.
This combination of population analysis, anatomical clustering and computational modeling strengthens the case for a modular organization in the brain. Neural activity is often described as highly distributed, with many regions contributing to behavior at once. The new findings do not reject that view; instead, they refine it. A distributed brain can still contain specialized communities of neurons whose activity is coordinated around particular computations. These communities may be flexible rather than rigid, allowing the same neural resources to be deployed in different situations. Such an arrangement would give the brain both efficiency and adaptability, two properties that are difficult to achieve simultaneously.
The study also offers a possible explanation for how animals perform complex tasks under changing conditions. If cognitive operations are built from reusable modules, learning a new task may involve modifying how modules are connected or coordinated rather than constructing entirely new circuits. This could reduce the biological cost of learning and help explain why the brain can rapidly generalize knowledge from one situation to another. The findings may also influence the design of artificial intelligence systems. Neural networks with reusable modules could potentially learn more efficiently, adapt to unfamiliar inputs and avoid duplicating the same computation across every task.
Although the work does not establish that all cognition is organized in discrete modules, it provides evidence that reusable computational components exist within cortical networks supporting working memory and stimulus processing. The results point toward a brain that is neither a collection of isolated specialists nor an undifferentiated web in which every neuron does everything. Instead, cognition may emerge from a dynamic compromise: distinct neuronal groups perform specialized operations, while those operations remain available for reuse in new combinations. For neuroscience, that idea offers a framework for studying flexibility at the level of neural populations. For artificial intelligence, it presents a biological example of how modular systems might produce adaptable, complex behavior.
Subject of Research: Reusable neuronal subspaces and modular brain architecture supporting flexible cognitive operations in mice and artificial recurrent neural networks.
Article Title: Reusable modular architecture enables flexible cognitive operations in the mouse brain and artificial recurrent networks
Article References: Osako, Y., Heller, G.R., Ährlund-Richter, S. et al. “Reusable modular architecture enables flexible cognitive operations in the mouse brain and artificial recurrent networks.” Nature Neuroscience (2026). https://doi.org/10.1038/s41593-026-02410-0
Image Credits: AI Generated
DOI: https://doi.org/10.1038/s41593-026-02410-0
Keywords: neuroscience, cognitive flexibility, modular brain architecture, neuronal subspaces, working memory, prefrontal cortex, parietal cortex, artificial recurrent networks, computational neuroscience, mouse brain
Tags: brain as a library of reusable computationscortical network reusabilitycortical organization in complex behaviorflexible cognition in artificial neural networksimplications for machine learning and AI designmodularity in biological and artificial intelligenceneural basis of memory and sensory processingneural circuit organization for adaptive behaviorneuroscience of flexible task performancereusable neural components in miceReversible brain modular architecturespecialized computational modules in prefrontal cortex
