Curiosity may be what drives our brains to become so complex

How did our brains evolve to be so complex? The answer may boil down to curiosity. Researchers who trained thousands of artificial neural networks with different learning styles found that the most curious ones not only had more sophisticated and flexible cognitive abilities, they also spontaneously organised into structures that resemble human brains. The networks formed interconnected modules and followed a human-like developmental trajectory, for example, with dense neural connections forming early on before being pruned away. Curiosity is usually seen as a consequence of having a complex brain, says Francesco Poli at the University of Cambridge. “What if it’s actually helping to shape how the brain is?” To model how infants learn, Poli and his colleagues collected eye-tracking data from 275 babies who were between 8 and 15 months old. The infants were shown sequences of colourful shapes on a screen, which provided varying amounts of information about where the next shape was likely to be displayed. “The positions weren’t random, and every different shape had a sequence that could be learned,” says Poli. Advertisement The babies proved to be highly sensitive to how much information the different stimuli provided. Whenever their rate of learning dropped below average, they disengaged and looked for something new, says Poli. The team used this principle to design artificial neural networks with different levels of intrinsic curiosity, then trained them on 20 tasks. The high-curiosity models were designed to immediately switch task if their rate of learning dropped below a certain threshold, whereas the low-curiosity models switched randomly. All the networks, curious or not, became experts at the tasks they were trained on. But after the training, the curious models performed significantly better at tests they hadn’t seen before. This ability to put together different things in flexible ways is seen as a hallmark of complex human cognition, says Poli. The curious models also developed physical characteristics resembling those of complex brains. The researchers embedded the artificial networks into a virtual three-dimensional space, modelled on the anatomy of the human brain. They then forced the networks to wire themselves efficiently by penalising long-distance connections, in order to mimic the cost of long-range wiring in real brains. On its own, this wasn’t enough to produce complexity, but when the researchers added curiosity on top, a range of brain-like features emerged. For example, the curious networks developed a hierarchical “modular” structure with highly-interconnected areas that were in turn connected to each other. Poli was also surprised to see a particular trajectory in the curious networks: the density of their connections rose, then fell. “I thought that was weird,” he says. But this is similar to what happens in human brains, says Poli, where the number of synapses – the junctions through which neurons communicate – rises steeply during the first year of life, followed by a pruning phase where unhelpful connections are lost. He speculates that curiosity, the drive to seek out tasks that yield the most information, could be “related to this selection process of what to keep and what not to keep”. To estimate a biological age for each network, the researchers mapped the rise and fall of the neural networks’ densities against a typical human profile. They found that the most curious, human-like ones were at a more mature phase in the developmental arc, with an age of around 20 years, whereas those that didn’t have the combination of curiosity and the correct wiring were stuck at an earlier point, with an age of around 9 years. It is unsurprising that curiosity makes learning more efficient, says Alireza Modirshanechi at the University of Göttingen in Germany, because it ensures we don’t waste time on tasks that are too easy or too hard. “You try to learn those that are most learnable,” he says. But the finding that this enables the emergence of human-like brain architecture is “very surprising”, says Modirshanechi. “If true, it’s a very strong claim.” He says it will be important to check that the non-curious networks don’t eventually develop similar characteristics, albeit more slowly, if they are given enough training time. Poli argues that the non-curious networks performed just as well as the curious ones on the basic training tasks, so their lack of complex cognition and brain-like features probably isn’t simply a question of learning slower. “I think that, even with more training time, they would have stayed pretty much the same,” he says. “Curiosity really changed how the networks learned.” Regardless, Poli says the findings emphasise the role of curiosity in education. Environments that force children to do one particular activity for a length of time “might be counterproductive, especially early on”, he says. “It depends on what we want. If we value the ability to stay still and work on one thing for a long amount of time, then we have to train that in kids.” Whereas if we value curiosity and cognition, we may need to prioritise giving children a choice of activities, “then letting them free within the space”, he says. The results also feed into debates around the development of flexible, human-like artificial intelligence. The vast majority of AI research “is just building bigger systems”, says Poli, that consume ever-greater data and resources. He argues that, for complex cognition, it might be fruitful to build curious AIs that learn more like people. The fact that evolution doesn’t have endless resources can be what drives clever solutions, he says. “Sometimes, these constraints are also what help us to excel.” Modirshanechi isn’t convinced that AI developers should necessarily switch tack. Large language models such as ChatGPT show that scaling works, he says. But incorporating principles of curiosity might lead to more human-compatible systems, he says, for example models that would ask enough questions to figure out users’ intentions before providing lengthy answers. bioRxiv DOI: 10.64898/2026.07.02.735826