Researchers have found that the brain repeatedly relies on the same cognitive “building blocks” when performing different types of tasks. By rewiring these blocks in new ways, the brain can quickly generate new behaviors.
Princeton researchers have found that the brain uses reusable “thinking blocks” to quickly create new behaviors.
Artificial intelligence can already produce acclaimed essays and support medical diagnoses with impressive accuracy. However, biological brains still outperform machines in one crucial area: flexibility. Humans absorb new information and adapt to unfamiliar situations with very little effort. People can log into a new program, follow a recipe they’ve never tried, or learn the rules of a game they’ve just discovered. In contrast, AI systems often struggle to adapt in real time and truly learn “on the fly.”
A new study by Princeton neuroscientists offers insight into why the brain excels at this kind of rapid adaptation. The researchers found that the brain repeatedly relies on the same cognitive “building blocks” when performing different types of tasks. By reassembling these blocks in new ways, the brain can quickly generate new behaviors.
“Advanced AI models can achieve human, and even superhuman, performance on a single task,” said Tim Buschman, senior author of the study and deputy director of the Princeton Institute for Neuroscience. “But they struggle to learn and perform many different tasks. We found that the brain is flexible because it can reuse thinking components across many different tasks. By ‘snapping’ these thinking blocks together, the brain is able to build new tasks.”
The study was published on November 26, 2025 in the journal Nature.
Compositionality: Building new skills from familiar ones
People often learn something new based on similar abilities they already have. Someone who knows how to maintain a bicycle, for example, may learn to repair a motorcycle relatively easily. Scientists call this process of building new skills from simple, existing skills compositionality.
“If you already know how to bake bread, you can use that ability to bake a cake without relearning how to bake from scratch,” said Sina Tafazoli, a postdoctoral fellow in Bushman’s lab and first author of the paper. “You’re repurposing existing skills—using an oven, measuring ingredients, kneading dough—and combining them with new skills, like whipping batter and making frosting, to create something completely different.”
Although compositionality is considered central to human flexibility, evidence about how the brain accomplishes this has been limited and sometimes inconsistent.
To test the idea more closely, Tapazzoli trained two male rhesus macaques to perform three related tasks, recording activity in different areas of their brains.
Video clip from a color-or-shape discrimination task presented to study subjects. Credit: Sina Tapazzoli (Princeton University).
How did monkeys reveal the brain's learning strategy?
Instead of “real-world” tasks like repairing cars or baking, the monkeys performed visual categorization challenges. They were shown colorful, balloon-like blobs on a screen and asked to decide whether each blob looked more like a rabbit or the letter T (shape categorization), or whether it looked more red or green (color categorization).
The level of difficulty varied greatly from trial to trial. Some images clearly resembled a rabbit or were painted a bright red, while others were vague and required much more careful judgment.
To report their choice, each monkey looked in one of four directions. In one task, looking left signaled “rabbit” and looking right signaled “T.”
A key element of the experimental design was that each task had its own rules, but also elements in common with the others. One of the color tasks and the shape task required the monkeys to look in the same directions to report a response, while both color tasks required the monkeys to categorize a color in the same way (more red or more green), although the gaze directions required for the report were different.
This structure allowed researchers to test whether the brain relied on the same patterns of neural activity—that is, cognitive “building blocks”—whenever tasks had overlapping components.
The prefrontal cortex stores reusable cognitive blocks
When Tapazzoli and Bushman analyzed brain activity, they found that the prefrontal cortex, a region involved in higher-level thinking, contained several repetitive patterns of neural activity. These patterns appeared across different tasks whenever neurons worked toward a common goal, such as distinguishing colors.
Bushman called these shared patterns “cognitive Lego”—a set of building blocks that can be assembled in different ways to produce new behaviors.
“I think of a cognitive block as a function in a computer program,” Bushman said. “One group of neurons might distinguish a color, and its output might be mapped to another function that drives an action. This organization allows the brain to perform a task by sequentially executing each component of the task.”
In one color task, for example, the brain combined a block that evaluated color with another block that directed eye movements in different directions. When the monkeys switched from color evaluation to shape recognition, using similar movements, the brain simply activated the shape-processing block along with the same eye-movement block.
This pattern of sharing was particularly strong in the prefrontal cortex, and appeared much less in other brain regions, suggesting that compositionality may be a unique role of the prefrontal cortex.
Decrease in activity of unnecessary blocks helps the brain stay focused
Tapazzoli and Bushman also found that the prefrontal cortex reduces the activity of certain cognitive blocks when they are not needed. This likely helps the brain focus more effectively on the most relevant task.
“The brain has a limited capacity for cognitive control,” Tapazzoli said. “You have to ‘compress’ some of your abilities to focus on what’s important at the moment. Focusing on categorizing shape, for example, temporarily reduces your ability to encode color, because the goal is to distinguish shape, not color.”
Selective activation and suppression of cognitive blocks may allow the brain to avoid overload and remain focused on the immediate goal.
What does “Cognitive Lego” say about artificial intelligence and human health?
Such a “Lego” may explain why humans learn new tasks so quickly. Instead of building each behavior from scratch, the brain reuses existing components and prevents duplication of work—something that today’s AI systems typically lack.
“A central problem in machine learning is catastrophic interference,” Tapazzoli said. “When a machine or a network of neurons learns something new, it forgets and erases previous memories. If an artificial neural network knows how to bake a cake and then learns to bake cookies, it will forget how to bake a cake.”
Incorporating compositionality into artificial intelligence may help develop systems that add new skills while maintaining repetition.
This insight may also provide direction for understanding neurological and psychiatric disorders. Conditions such as schizophrenia, obsessive-compulsive disorder, and certain types of brain injury can make it difficult for people to apply familiar skills in new contexts. These difficulties may stem from disruptions in the brain's ability to integrate and reuse its cognitive building blocks.
“Imagine we could help people regain the ability to switch strategies, learn new routines, or adapt to change,” Tapazzoli said. “In the long term, understanding how the brain reuses and recomposes knowledge could help us design treatments that restore this process.”
Scientific reference (for the scientific article):
"Building compositional tasks with shared neural subspaces” by Sina Tafazoli, Flora M. Bouchacourt, Adel Ardalan, Nikola T. Markov, Motoaki Uchimura, Marcelo G. Mattar, Nathaniel D. Daw and Timothy J. Buschman, 26 November 2025, Nature.
DOI: 10.1038/s41586-025-09805-2
More of the topic in Hayadan:
4 תגובות
This is just a small explanation for the inferiority of the royal intellect compared to the brain. Even in a new language learning task, every baby manages to learn a language almost perfectly within a few years (without relying on existing cognitive blocks). A language model cannot learn a language without an amount of text equivalent to hundreds of years of listening to the language.
But of course, the machine, as long as the task is well defined, has other advantages.
Just as no engine that humans have ever built will ever reach the level of energy utilization of the human digestive system and muscles, yet as long as the task is to move forward on relatively straight terrain, a car will easily catch up with the muscles.
What is the difference between machine intelligence and human intelligence?
Human intelligence has natural knowledge, and machine intelligence does not have natural knowledge.
What is natural knowledge?
To a person of natural intelligence who touches a block of ice, comes a wondrous knowledge of feeling,
A machine with machine intelligence touching a block of ice does not receive any miraculous knowledge of feeling.
The denominator does not feel cold, and the person knows the wonderful feeling of cold because he has natural knowledge.
This explanation is based on the assumption that ancient man gave a name to a miraculous message that came to man following contact with a block of ice, and this name consists of two sounds that we say today, resulting from the letters ككككككك رككككك ركككك
Human language is a language of noises, based on human natural knowledge, and bird language is also a language of noises based on the natural knowledge of birds,
Every noise expresses a natural knowledge that follows an action, and this is true for all living beings.
Man invented a number of noises and gave them the name letters, and this is the secret of human language.
The language of living beings is a language of noises based on their natural knowledge,
Therefore, noise cannot be explained by other noise.
Artificial Intelligence (Ai) is just a robot
Who plays with algorithms and repeats himself, until boredom and complications arise.
His, all the paintings are generally as clean as a rubber band from Mort.
As a painter, I make a lot of mistakes, so there are a lot of corrections that the AI is not capable of.
To do the work of a true artist
It's true, he serves as an excellent guide for me and every question is answered with an answer.
Sometimes he mixes things up and I have to point it out to him, just like that.
However, he is a pretty pleasant and good teacher with a lot of knowledge.
But sometimes when he is given an assignment to write a letter, he sometimes adds a lot of irrelevant things.
AI is basically useful
But he is not the monster they make him out to be.
Yes, but artificial intelligence can overcome shortcuts by considering millions of possibilities that the brain cannot do. In doing so, it can detect things that the brain with shortcuts did not find or will not find.