The brain remixes what it already knows to solve problems it has never seen
In a laboratory at Rockefeller University, macaques solved drawing problems they had never encountered before—not through trial and error, but by silently recombining movements they already knew. Neuroscientists, recording from eight brain regions, found that this capacity for compositional generalization is housed in the ventral premotor cortex, a region long associated with movement rather than abstract thought. The discovery suggests that creativity—the act of making something new from something known—may have a precise, mappable address in the brain, and that the line between motor habit and symbolic reasoning is thinner than we imagined.
- Macaques drew complex, never-before-seen shapes correctly on their very first attempt, with no practice or instruction—a result that stopped researchers in their tracks.
- The finding directly challenges the prevailing assumption that the prefrontal cortex governs abstract rule-following, relocating that function to a motor-planning region instead.
- Electrode arrays tracking 40–50 neurons across eight brain regions revealed that only the ventral premotor cortex satisfied all three criteria for compositional, symbol-based encoding.
- Each monkey solved the novel tasks in its own distinctive style, suggesting that learned 'action symbols' are personal and flexible rather than rigid or universal.
- The team is now tracing how these symbols organize into grammar-like hierarchies, with early evidence pointing to the pre-supplementary motor area as a kind of syntactic coordinator.
A macaque traces a circle on a touch screen. Another draws a zigzag. A third, a chevron. These are not mere tricks—they are the building blocks of a discovery about how minds construct the new from the familiar.
When researchers at Rockefeller University later presented these same monkeys with complex shapes made by combining their learned symbols, the animals drew them correctly on the first try, without practice or instruction. Each monkey did so in its own distinctive way, as if signing its name. The phenomenon at work—using existing knowledge to solve novel problems—is called compositional generalization. Humans do it constantly, from poetry to jazz to mathematics. What science lacked was a clear picture of where, precisely, it happens in the brain.
To find out, the team implanted electrode arrays in two macaques, recording from 40 to 50 neurons across eight brain regions involved in planning, cognition, and motor control. They searched for three signatures: neural activity indifferent to a shape's size or position, distinct patterns for each simple symbol, and complex-shape responses that were simply the sum of their component parts.
Only one region satisfied all three: the ventral premotor cortex, a motor-planning area tucked into the lower front of the brain. There, the researchers identified what they called 'action symbols'—neural populations encoding learned movements in a form flexible enough to be recombined on demand. The result surprised even outside observers, since prevailing theory had assigned abstract rule-handling to the prefrontal cortex, a different region entirely.
The implications extend well beyond motor control. If the brain encodes actions as discrete, recombinable units—operating something like a symbolic computer program—then creativity and problem-solving may be mechanistically understandable rather than mysterious. The team is now mapping how these symbols arrange into grammar-like structures, with early findings implicating the pre-supplementary motor area in how symbols are ordered and combined.
The macaques keep drawing at their touch screens, solving problems on the first try. What was once opaque about the inventive mind is slowly, neuron by neuron, coming into focus.
A macaque sits before a touch screen, its finger tracing a circle in a single fluid stroke. Another draws a zigzag. A third makes a chevron. These are not tricks performed for treats, though juice rewards follow each success. They are the building blocks of something larger: evidence that brains—at least primate brains—solve novel problems by remixing what they already know.
When researchers later asked these same monkeys to draw complex shapes made from combinations of those learned symbols, something remarkable happened. The animals drew them correctly on the first attempt, without practice, without instruction. They had never seen these composite shapes before. Yet they succeeded by reaching into their personal repertoire of learned movements and assembling them in new ways—each monkey using its own distinctive style, like a handwriting all its own.
This capacity to solve problems by combining existing knowledge is not new to neuroscience. Humans do it constantly: a poet draws on learned words to write an opening line never written before; a jazz musician improvises by recombining familiar riffs; a mathematician solves an equation by assembling known functions in a novel order. The phenomenon has a name: compositional generalization. What has been missing, until now, is a clear picture of where in the brain this actually happens.
A team led by researchers at Rockefeller University implanted electrode arrays in the brains of two macaques, recording activity from 40 to 50 neurons across eight different brain regions known to be involved in planning, cognition, and motor control. As the monkeys learned their simple symbols and later attempted the complex shapes, the researchers watched the neural activity unfold. They were looking for three specific signatures: activity that would not change based on the size or position of a shape, since the brain should care only about the abstract structure of the movement, not its physical details; distinct neural patterns for each simple symbol; and, crucially, neural activity in response to complex shapes that was simply the sum of the activity for each component symbol.
Only one region showed all three signatures: the ventral premotor cortex, a brain area tucked into the lower front of the brain involved in planning and executing movements. There, researchers found what they called "action symbols"—neural populations that encoded the learned movements in a way that could be flexibly recombined. The finding was surprising enough to catch the attention of researchers outside the study. Previous work had suggested that the prefrontal cortex, a different region entirely, handled abstract rules and their implementation. Yet here was the ventral premotor cortex, doing something that looked very much like abstract symbolic reasoning, but localized to a single, specific area rather than distributed across the brain.
The implications ripple outward. If the brain encodes learned actions as discrete, recombinable units—if it works, in some sense, like a symbolic computer program with elementary components that can be assembled into larger structures—then creativity and problem-solving may not be mysterious or magical. They may be mechanistically understandable. The team is now exploring how the brain organizes these symbols into something resembling grammar, with loops and clauses and hierarchical structure, much as a programmer would write code. Early results suggest that different brain regions handle different aspects of this symbolic organization, with the pre-supplementary motor area in the dorsomedial frontal cortex playing a role in how symbols are arranged and combined.
For now, the macaques continue their work at the touch screen, drawing shapes that no monkey has drawn before, solving problems on the first try, their brains quietly performing the work of invention. What was once opaque—how a mind takes what it knows and builds something new from it—is beginning to come into focus.
Notable Quotes
We have quite a lot of behavioral evidence for compositional generalization across a wide array of different tasks. The interesting element here is the step towards showing a neural basis for that.— Charlie Wilson, INSERM and Stem Cell and Brain Research Institute
We're beginning to see some components, like loops and clauses, and elementary symbols that you would have in a computer program, in different parts of the brain, which suggests that the brain's inventiveness can be understood mechanistically.— Winrich Freiwald, Rockefeller University
The Hearth Conversation Another angle on the story
Why does it matter that this happens in the ventral premotor cortex and not somewhere else?
Because it tells us the brain isn't solving this problem by distributing the work everywhere at once. It's localized. That makes it tractable—we can actually study it, understand it, maybe even predict how it works.
The monkeys had never seen these complex shapes before, yet they drew them correctly on the first try. How is that possible?
They weren't drawing from memory of the shape itself. They were drawing from memory of the movements. The brain recognized that the new shape was made of pieces it already knew how to make, and it assembled those pieces in the right order.
So the brain is like a computer program?
In some ways, yes. It has elementary symbols—the simple movements—and it combines them according to rules, almost like syntax. But it's not a computer program. It's a brain. The analogy helps us understand it, but the actual mechanism is still being uncovered.
What happens next in this research?
They want to map the entire process—not just where the symbols are stored, but how the brain decides which symbols to use, in what order, and how it organizes them into something coherent. They're looking for the grammar of action.
Does this apply to humans?
That's the assumption, but it hasn't been proven yet. We see the same brain structures in humans. We know humans do compositional generalization all the time. But the direct evidence, the way they have it for macaques, doesn't exist yet.