If machines can learn how we reason, they may also help us understand why we do.
Centaur is a new artificial intelligence model designed to mimic human reasoning. Developed on Meta’s Llama 3.1 and trained on more than 10 million decisions from 160 psychology studies, it tackles logic problems, moral dilemmas and everyday choices to uncover the patterns behind human thinking. Researchers view it as a tool for discovery, one that behaves enough like a person to surface insights that traditional models might miss.
“I am excited about using the model to learn something about the human mind,” Marcel Binz, Deputy Head of the Institute for Human-Centered AI at Helmholtz Munich and lead author of the study describing Centaur, said in an interview with IBM Think.
A behavioral mirror for cognitive science
Most cognitive models strip experiments down to raw numbers. Centaur does the opposite. It reads each task in full, complete with natural language instructions and every step of the human response. The model was trained on a dataset called Psych 101, a collection of classic psychology problems that includes everything from visual puzzles and memory tests to moral dilemmas and language games. By seeing the same information a person would, Centaur learns to follow the task like a human.
That approach enabled generalization well beyond the training data. When researchers reworded a standard reinforcement learning problem, switching the framing from astronauts to magic carpets, Centaur still exhibited the same behavioral tendencies. It also performed well on entirely new types of tasks, such as LSAT-style logic puzzles.
The use of language, rather than compressed numerical descriptions, was deliberate. “We wanted the model to see what participants saw,” Binz explained. “Full instructions, full context. No shortcuts.”
Centaur isn’t built to explain the workings of the brain. Instead, it focuses on reproducing what people do in behavioral studies. That predictive power has immediate implications for researchers, who often rely on narrow, hand-built models for each type of cognitive function.
Russell Poldrack, a Professor of Psychology at Stanford University who was not involved in the project, views Centaur as part of a larger shift in the field.
“Historically, we’ve given models highly reduced versions of tasks,” he told IBM Think in an interview. “Now, we can give them what we’d give a person and see behavior that mirrors what a person would do.”
The difference isn’t just in scale, but in intent. Most cognitive models are constructed to explain a specific behavior. Centaur is built to observe and replicate behavior across domains, such as visual reasoning and memory tasks. That opens the possibility of discovering new patterns that researchers might otherwise miss.
In one example from the study, the team examined how people choose between products with multiple expert ratings. Centaur’s behavior revealed a two-step strategy: people initially appeared to count the number of positive ratings, and only used expert credibility as a tiebreaker. That insight led to a new, interpretable model of human decision-making, one that Centaur was able to match after refinement.
“We’re not trying to replace cognitive models,” said Binz. “We want to give researchers better tools for exploring what people might be doing.”
Looking for failure points
Even with its breadth, Centaur has well-defined limits. It doesn’t simulate timing, attention dynamics or physical interaction. It can’t explain how long a person takes to respond, or how behavior changes in a social setting or how decisions unfold over time.
Those limits may prove useful. Where Centaur fails, researchers may find clues about the aspects of cognition that aren’t easily learned from language alone.
That’s exactly where Poldrack would start. “I’d want to go find the places where it breaks,” he said. “What does it miss? Where does it diverge from what people do—and why?”
Centaur’s architecture, a type of transformer, is not designed to model complex cognitive dynamics. Recurrence, memory modules or multimodal training may be needed to bring it closer to those capabilities. But even now, its ability to produce human-like behavior across a wide set of tasks is unusual.
Memory, not magic
Some researchers have questioned whether large language models (LLMs) are truly reasoning at all, or whether they are merely repeating what they’ve seen during training. Binz chooses his words carefully when he describes Centaur. “It’s not simulating how a human brain works,” he said. “But it’s also not just copying. It’s doing something that generalizes.”
Some researchers have questioned whether large language models (LLMs) are truly reasoning or simply repeating patterns they encountered during training. Binz chooses his words carefully when he describes Centaur. “It’s not simulating how a human brain works,” he said. “But it’s also not just copying. It’s doing something that generalizes.”
Poldrack noted that this debate isn’t new. The question of whether language models are genuinely thinking or merely imitating statistical patterns in language has been at the center of AI criticism, often referred to as the “stochastic parrot” problem.
“When people first started throwing the stochastic parrot criticism at large language models, my initial response was that it’s pretty clear humans are at least somewhat stochastic parrots, too,” he said.
He pointed to exemplar theory, a concept from psychology that holds that people often rely on specific memories rather than abstract rules when making decisions.
“When I see a dog, I’m not recomputing what a dog is,” he said. “I’m matching it to something I’ve seen before. That’s fast and it works.”
Poldrack suggested that Centaur might be drawing on past experiences, combining them in new ways and generating predictions. But whether this process amounts to thinking remains an open question, he said.
Data as a new foundation
One of the reasons models like Centaur are possible now is that the data has finally caught up with the questions. For decades, psychology operated in what Poldrack described as a “data-limited regime,” with experiments involving 30 or 40 participants, analyzed by hand.
Psych-101 changes that. The dataset brings together tens of millions of decisions drawn from decades of psychology research, all rewritten in a consistent, natural language format. It includes complete task descriptions, instructions and sequences of human responses across a wide range of experiments. This is the foundation on which Centaur was trained. Instead of learning from isolated inputs and outputs, the model is exposed to the full context of each task. That allows it to engage with problems in a more human-like way, following the structure and flow of each experience.
That scale may not provide deep explanations, Poldrack said, but it opens new doors for exploration.
“We’ve never had access to this kind of data regime before,” he said. “Now we can train models that reflect behavior across tasks, not just within them.”
Binz said the team plans to expand Psych-101 in the coming months to include psycholinguistics, developmental studies and cross-cultural tasks. The goal is to do more than match average behavior. Researchers want to understand how people differ based on age, personality or background, and how those differences shape the way they respond.
“Eventually, we want to build models that can reason about cognition itself,” he said.
Centaur doesn’t pretend to be a brain. But it may be something else cognitive science has lacked: a general-purpose behavioral model, trained at scale, that comports itself similarly enough to a person to help us understand where our theories succeed, and where they don’t.
“It’s essentially a big black box that predicts behavior really well,” Binz said. “But the more we understand what’s inside, the more we may learn about what’s inside us, too.”
IBM News