Centaur AI: Revolutionizing Human Behavior Prediction and Cognitive Science

This AI 'thinks' like a human -- after training on just 160 psychology studies

An innovative artificial-intelligence (AI) system can predict the decisions people will make in a wide variety of situations -- often outperforming classical theories used in psychology to describe human choices. The researchers who developed the system, called Centaur, fine-tuned a large language model (LLM) using a massive set of data from 160 psychology experiments, in which 60,000 people made more than 10 million choices across many tasks. Most computer models and cognitive theories stick to a single task. For instance, Google Deepmind's AlphaGo can only play the strategy game Go, and prospect theory can only predict how a person will choose between potential losses and gains. Centaur, by contrast, can simulate human behaviour across a spectrum of tasks -- including gambling, memory games and problem-solving. During testing, it was even able to predict people's choices in tasks it had not been trained on. The development of Centaur is described in a paper published today in Nature. The team that created the system thinks that it could one day become a valuable tool in cognitive science. "You can basically run experimental sessions in silico instead of running them on actual human participants," says study co-author Marcel Binz, a cognitive scientist at the Helmholtz Institute for Human-Centered AI in Munich, Germany. That could be useful when conventional studies would be too slow, he says, or when it's difficult to recruit children or people with psychiatric conditions. "Building theories in cognitive science is very difficult," says Giosuè Baggio, a psycholinguist at the Norwegian University of Science and Technology in Trondheim, Norway. "It's exciting to see what we can come up with with help from machines." Scientists have long struggled with using task-specific models to simulate broad aspects of human behaviour because the tools cannot generalize to a multitude of tasks. Binz and his colleagues wanted to overcome this limitation. They spent five days fine-tuning Llama -- an LLM released by the technology firm Meta in Menlo Park, California -- with a huge behavioural data set called 'Psych 101'. The researchers tuned the model to predict not just an average behaviour for a given task, but the range of typical behaviours in the population. They then tested how well it predicted the behaviour of participants who were not included in the training data set. In all but 1 of 32 tasks, Centaur outperformed Llama and 14 cognitive and statistical models in predicting what choices participants would make. That one outlier was a task in which participants judged whether sentences were grammatically correct. Centaur also performed well when given altered versions of the tasks that it was trained on, and even tasks that didn't resemble any in its training, such as logical reasoning. This "shows that there's a lot of structure in human behaviour", says Russell Poldrack, a cognitive neuroscientist at Stanford University in California. "It really ups the bar for the power of the kinds of models that psychology should be aspiring to." Despite its broad capabilities, researchers say that Centaur still has limitations. The AI model relies entirely on language-based tasks, Poldrack says. For example, it can predict what a person might choose during a given task, but it "can't predict how long it's going to take them" to make their choice. The authors say that they are expanding the training data set so that it is up to four times as large as the existing one. Many of the current data come from Western, educated, industrialized populations, which the researchers acknowledge might limit how well Centaur's predictions apply to diverse groups. Binz says that the next step for Centaur, which is freely accessible, is that it "needs to be externally validated by the research community". "Right now," he adds, "it's probably the worst version of Centaur that we will ever have, and it will only get better from here."

How scientists are trying to use AI to unlock the human mind

In one of the studies, researchers transformed a large language model into what they refer to as a "foundation model of human cognition." Out of the box, large language models aren't great at mimicking human behavior -- they behave logically in settings where humans abandon reason, such as casinos. So the researchers fine-tuned Llama 3.1, one of Meta's open-source LLMs, on data from a range of 160 psychology experiments, which involved tasks like choosing from a set of "slot machines" to get the maximum payout or remembering sequences of letters. They called the resulting model Centaur. Compared with conventional psychological models, which use simple math equations, Centaur did a far better job of predicting behavior. Accurate predictions of how humans respond in psychology experiments are valuable in and of themselves: For example, scientists could use Centaur to pilot their experiments on a computer before recruiting, and paying, human participants. In their paper, however, the researchers propose that Centaur could be more than just a prediction machine. By interrogating the mechanisms that allow Centaur to effectively replicate human behavior, they argue, scientists could develop new theories about the inner workings of the mind. But some psychologists doubt whether Centaur can tell us much about the mind at all. Sure, it's better than conventional psychological models at predicting how humans behave -- but it also has a billion times more parameters. And just because a model behaves like a human on the outside doesn't mean that it functions like one on the inside. Olivia Guest, an assistant professor of computational cognitive science at Radboud University in the Netherlands, compares Centaur to a calculator, which can effectively predict the response a math whiz will give when asked to add two numbers. "I don't know what you would learn about human addition by studying a calculator," she says. Even if Centaur does capture something important about human psychology, scientists may struggle to extract any insight from the model's millions of neurons. Though AI researchers are working hard to figure out how large language models work, they've barely managed to crack open the black box. Understanding an enormous neural-network model of the human mind may not prove much easier than understanding the thing itself. One alternative approach is to go small. The second of the two Nature studies focuses on minuscule neural networks -- some containing only a single neuron -- that nevertheless can predict behavior in mice, rats, monkeys, and even humans. Because the networks are so small, it's possible to track the activity of each individual neuron and use that data to figure out how the network is producing its behavioral predictions. And while there's no guarantee that these models function like the brains they were trained to mimic, they can, at the very least, generate testable hypotheses about human and animal cognition. There's a cost to comprehensibility. Unlike Centaur, which was trained to mimic human behavior in dozens of different tasks, each tiny network can only predict behavior in one specific task. One network, for example, is specialized for making predictions about how people choose among different slot machines. "If the behavior is really complex, you need a large network," says Marcelo Mattar, an assistant professor of psychology and neural science at New York University who led the tiny-network study and also contributed to Centaur. "The compromise, of course, is that now understanding it is very, very difficult." This trade-off between prediction and understanding is a key feature of neural-network-driven science. (I also happen to be writing a book about it.) Studies like Mattar's are making some progress toward closing that gap -- as tiny as his networks are, they can predict behavior more accurately than traditional psychological models. So is the research into LLM interpretability happening at places like Anthropic. For now, however, our understanding of complex systems -- from humans to climate systems to proteins -- is lagging farther and farther behind our ability to make predictions about them.

Scientist Use A.I. To Mimic the Mind, Warts and All

To better understand human cognition, scientists trained a large language model on 10 million psychology experiment questions. It now answers questions much like we do. Companies like OpenAI and Meta are in a race to make something they like to call artificial general intelligence. But for all the money being spent on it, A.G.I. has no settled definition. It's more of an aspiration to create something indistinguishable from the human mind. Artificial intelligence today is already doing a lot of things that were once limited to human minds -- such as playing championship chess and figuring out the structure of proteins. ChatGPT and other chatbots are crafting language so humanlike that people are falling in love with them. But for now, artificial intelligence remains very distinguishable from the human kind. Many A.I. systems are good at one thing and one thing only. A grandmaster can drive a car to a chess tournament, but a chess-playing A.I. system is helpless behind the wheel. An A.I. chatbot can sometimes make very simple -- and very weird -- mistakes, like letting pawns move sideways in chess, an illegal move. For all these shortcomings, an international team of scientists believe that A.I. systems can help them understand how the human mind works. They have created a ChatGPT-like system that can play the part of a human in a psychological experiment and behave as if it has a human mind. Details about the system, known as Centaur, were published on Wednesday in the journal Nature. In recent decades, cognitive scientists have created sophisticated theories to explain various things that our minds can do: learn, recall memories, make decisions and more. To test these theories, cognitive scientists run experiments to see if human behavior matches a theory's predictions. Some theories have fared well on such tests, and can even explain the mind's quirks. We generally choose certainty over risk, for instance, even if that means forgoing a chance to make big gains. If people are offered $1,000, they will usually take that firm offer rather than make a bet that might, or might not, deliver a much bigger payout. But each of these theories tackles only one feature of the mind. "Ultimately, we want to understand the human mind as a whole and see how these things are all connected," said Marcel Binz, a cognitive scientist at Helmholtz Munich, a German research center, and an author of the new study. Three years ago, Dr. Binz became intrigued by ChatGPT and similar A.I. systems, known as large language models. "They had this very humanlike characteristic that you could ask them about anything, and they would do something sensible," Dr. Binz said. "It was the first computation system that had a tiny bit of this humanlike generality." At first Dr. Binz could only play with large language models, because their creators kept the code locked away. But in 2023, Meta released the open-source LLaMA (Large Language Model Meta AI). Scientists could download and modify it for their own research. (Thirteen authors have sued Meta for copyright infringement, and The New York Times has sued OpenAI, ChatGPT's creator, and its partner, Microsoft.) The humanlike generality of LLaMA led Dr. Binz and his colleagues to wonder if they could train it to behave like a human mind -- not just in one way but in many ways. For this new lesson, the scientists would present LLaMA with the results of psychological experiments. The researchers gathered a range of studies to train LLaMA -- some that they had carried out themselves, and others that were conducted by other groups. In one study, human volunteers played a game in which they steered a spaceship in search of treasure. In another, they memorized lists of words. In yet another, they played a pair of slot machines with different payouts and figured out how to win as much money as possible. All told, 160 experiments were chosen for LLaMA to train on, including over 10 million responses from more than 60,000 volunteers. Dr. Binz and his colleagues then prompted LLaMA to play the part of a volunteer in each experiment. They rewarded the A.I. system when it responded in a way that a human had. "We essentially taught it to mimic the choices that were made by the human participants," Dr. Binz said. He and his colleagues named the modified model Centaur, in honor of the mythological creature with the upper body of a human and the legs of a horse. Once they trained Centaur, the researchers tested how well it had mimicked human psychology. In one set of trials, they showed Centaur some of the volunteer responses that it hadn't seen before. Centaur did a good job of predicting what a volunteer's remaining responses would look like. The researchers also let Centaur play some of the games on its own, such as using a spaceship to find treasure. Centaur developed the same search strategies that human volunteers had figured out. To see just how humanlike Centaur had become, the scientists then gave it new games to play. In the spaceship experiment, scientists had changed the story of the game, so that now volunteers rode a flying carpet. The volunteers simply transferred their spaceship strategy to the new game. When Dr. Binz and his colleagues made the same switch for Centaur, it transferred its spaceship strategy, too. "There is quite a bit of generalization happening," Dr. Binz said. The researchers then had Centaur respond to logical reasoning questions, a challenge that was not in the original training. Centaur once again produced humanlike answers. It tended to correctly answer questions that people got right, and failed on the ones that people likewise found hard. Another human quirk emerged when Dr. Binz and his colleagues replayed a 2022 experiment that explored how people learn about other people's behavior. In that study, volunteers observed the moves made by two opposing players in games similar to Rock, Paper, Scissors. The observers figured out the different strategies that people used and could even predict their next moves. But when the scientists instead generated the moves from a statistical equation, the human observers struggled to work out the artificial strategy. "We found that was exactly the same case for Centaur as well," Dr. Binz said. "The fact that it actually predicts the human players better than the artificial players really means that it has picked up on some kind of things that are important for human cognition." Some experts gave Centaur high marks. "It's pretty impressive," said Russ Poldrack, a cognitive scientist at Stanford University who was not involved in the study. "This is really the first model that can do all these types of tasks in a way that's just like a human subject." Ilia Sucholutsky, a computer scientist at New York University, was struck by how well Centaur performed. "Centaur does significantly better than classical cognitive models," he said. But other scientists were less impressed. Olivia Guest, a computational cognitive scientist at Radboud University in the Netherlands, argued that because the scientists hadn't used a theory about cognition in building Centaur, its prediction didn't have much to reveal about how the mind works. "Prediction in this context is a red herring," she said. Gary Lupyan, a cognitive scientist at Indiana University, said that theories that can explain the mind are what he and his fellow cognitive scientists are ultimately chasing. "The goal is not prediction," he said. "The goal is understanding." Dr. Binz readily agreed that the system did not yet point to a new theory of the mind. "Centaur doesn't really do that yet, at least not out of the box," he said. But he hopes that the language model can serve as a benchmark for new theories, and can show how well a single model can mimic so many kinds of human behavior. And Dr. Binz hopes to expand Centaur's reach. He and his colleagues are in the process of increasing their database of psychological experiments by a factor of five, and they plan on training the system further. "I would expect with that data set, you can do even more stuff," he predicted.

This AI model was trained on 10M human choices. Now it thinks and reacts like us

Named Centaur, in honor of the mythological creature with the upper body of a human and the legs of a horse, the AI model is trained on more than ten million individual decisions made by more than 60,000 participants across 160 psychological experiments. It reflects how people think and make choices in both familiar and novel situations, opening new avenues for understanding human cognition and improving psychological theories. For decades, cognitive science has struggled to build models that can both explain and predict human behavior. Traditional approaches have typically managed one or the other, but not both. Centaur changes that. Developed by Dr. Marcel Binz and Dr. Eric Schulz at the Institute for Human-Centered AI, the model uses a curated dataset called Psych-101, capturing a wide range of behavioral patterns, including, risk-taking reward learning and moral reasoning, and structures them in a format a language model can understand. Unlike earlier models that rely on predefined rules or narrow parameters, Centaur learns common decision-making strategies and generalizes them to new contexts. It even predicts reaction times, offering a deeper view into the dynamics of choice.

Centaur: AI that thinks like us -- and could help explain how we think

Researchers at Helmholtz Munich have developed an artificial intelligence model that can simulate human behavior with remarkable accuracy. The language model, called Centaur, was trained on more than ten million decisions from psychological experiments -- and makes decisions in ways that closely resemble those of real people. This opens new avenues for understanding human cognition and improving psychological theories. For decades, psychology has aspired to explain the full complexity of human thought. Yet traditional models could either offer a transparent explanation of how people think -- or reliably predict how they behave. Achieving both has long seemed out of reach. The team led by Dr. Marcel Binz and Dr. Eric Schulz, both researchers at the Institute for Human-Centered AI at Helmholtz Munich, has now developed a model that combines both. Centaur was trained using a specially curated dataset called Psych-101, which includes over ten million individual decisions from 160 behavioral experiments. The study is published in the journal Nature. What makes Centaur unique is its ability to predict human behavior not only in familiar tasks, but also in entirely new situations it has never encountered before. It identifies common decision-making strategies, adapts flexibly to changing contexts -- and even predicts reaction times with surprising precision. "We've created a tool that allows us to predict human behavior in any situation described in natural language -- like a virtual laboratory," says Binz, who is also the study's lead author. Potential applications range from analyzing classic psychological experiments to simulating individual decision-making processes in clinical contexts -- for example, in depression or anxiety disorders. The model opens up new perspectives in health research in particular -- for example, by helping us understand how people with different psychological conditions make decisions. The dataset is set to be expanded to include demographic and psychological characteristics. Centaur: Bridging theory and prediction Centaur bridges two previously separate domains: interpretable theories and predictive power. It can reveal where classical models fall short -- and provide insights into how they might be improved. This opens up new possibilities for research and real‑world applications, from medicine to environmental science and the social sciences. "We're just getting started and already seeing enormous potential," says institute director Schulz. Ensuring that such systems remain transparent and controllable is key, Binz adds -- for example, by using open, locally hosted models that safeguard full data sovereignty. Next, the researchers aim to take a closer look inside Centaur: Which computational patterns correspond to specific decision‑making processes? Can they be used to infer how people process information -- or how decision strategies differ between healthy individuals and those with mental health conditions? The researchers are convinced: "These models have the potential to fundamentally deepen our understanding of human cognition -- provided we use them responsibly." That this research is taking place at Helmholtz Munich rather than in the development departments of major tech companies is no coincidence. "We combine AI research with psychological theory -- and with a clear ethical commitment," says Binz. "In a public research environment, we have the freedom to pursue fundamental cognitive questions that are often not the focus in industry."

What is Centaur: AI that mimics human mind, say scientists

With brain-like internal patterns, Centaur may be the closest AI yet to human cognition In a major step forward for cognitive science and artificial intelligence, researchers at Helmholtz Munich have introduced Centaur, an AI system that can simulate and predict human behavior with a level of accuracy previously thought impossible. Developed by Dr. Marcel Binz and Dr. Eric Schulz at the Institute for Human-Centered AI, Centaur is being described as a "virtual laboratory for the mind." Built on Meta's Llama 3.1 language model and fine-tuned on a massive dataset of human decisions, it offers new ways to study cognition and raises important questions about how closely machines might one day resemble the human mind. Also read: Meta and Mark Zuckerberg bet big on AI Superintelligence: Here's how Psychologists have long wanted a tool that could do two things at once: explain how people think, and accurately predict how they'll behave. Existing models were often limited to one or the other. Centaur changes that. At its core is Psych-101, a dataset comprising over 10 million decisions from more than 60,000 people across 160 psychological experiments. These range from simple memory tasks to complex moral dilemmas. Every experiment was carefully standardized and rewritten in natural language so the AI could understand and learn from them. The result is an AI model that doesn't just reflect past human choices, it can anticipate new ones. Also read: Community Notes to fact-check: X becoming a playground for AI? What sets Centaur apart is its ability to generalize. It can handle entirely new situations and still make surprisingly accurate predictions about how a person would behave. Change the scenario, swap the setting, tweak the instructions and Centaur adapts. In tests, it outperformed long-standing cognitive models, including those specifically designed to study individual tasks. It even predicted how long someone might take to make a decision, showing a level of insight into human behavior that goes beyond language. One of Centaur's most interesting discoveries came when researchers compared its internal representations to human brain activity. Without ever being trained on neural data, the AI's behavior aligned with actual fMRI scans of people doing the same tasks. This suggests that in learning to model our decisions, Centaur developed a way of thinking that closely mirrors the computational structure of the human brain. Not just mimicking behavior, but echoing the way we process information. Researchers believe Centaur could play a major role in fields far beyond psychology. In medicine, it could help model how patients with anxiety or depression make decisions. In education, it could inform personalized learning systems. In design and marketing, it might help companies understand how users would react to a new product or interface, before it ever hits the market. Both Centaur and the Psych-101 dataset have been made publicly available to encourage collaboration and transparency. The research team is already working to expand the dataset to include more diverse populations and broader psychological domains. With this kind of predictive power, ethical considerations are front and center. The ability to model and anticipate human thought raises concerns around privacy, manipulation, and bias. The researchers have emphasized responsible use, hosting the model locally and promoting data sovereignty to protect user information. They're also clear-eyed about Centaur's limitations. Its current strengths lie in learning and decision-making, but less so in areas like social behavior or cultural variability. And like much of psychology, the data it's trained on is skewed toward Western, educated populations. Centaur doesn't offer a complete model of the human mind, but it brings researchers closer to that goal than ever before. It's not just a powerful AI tool, it's a system that reflects how we think, how we choose, and how we respond to the world around us. As it evolves, it could reshape not just how we study human behavior, but how we understand it.