A foundation model to predict and capture human cognition

0:00Welcome back to the deep dive. Here we sift through your source material, really digging deep to find those valuable insights, leaving you truly well informed, you know, without all the noise. Today, we're jumping into something pretty fascinating right at the heart of human cognition. It's all based on this incredible new paper from Nature, came out online July 2, 2025. And it's titled A Foundation Model to Predict and Capture Human Cognition. Now, you've probably heard about AI models getting really good at specific things, right, like mastering Go or writing text. But what if you could build just one single model, one that understands and predicts, well, the whole breadth of human intelligence, the general stuff?

0:40That's basically the huge ambition we're exploring today. Yeah. And what's really exciting, I think, is how this connects to a really longstanding goal in psychology. People have talked about a unified theory of cognition for ages. But most computational models up until now have been very domain specific. Meaning good at one thing. Exactly. Think AlphaGo. Brilliant at Go, right? But it can't pick out your breakfast cereal. And it's the same in cognitive science. We have models like prospect theory, great for decisions under risk, but they tell you nothing about learning or memory or exploration.

1:12Separate silos, basically. Okay. So we're really talking about a potential leap here, moving from these specialized tools towards something much bigger, a general purpose model of how we actually think and act. All right, let's unpack this. The paper introduces something called Centaur, and it's described, as you said, as a foundation model of human cognition. How on earth did they even start building something like that? So ambitious. Well, that's the core question, isn't it? How do you move towards a unified theory? Their approach was fundamentally data-driven. They took a cutting-edge large language model, specifically LAMA 3.170, a really powerful one, And they fine-tuned it, but not on text.

1:51They fine-tuned it on this absolutely enormous data set of actual human behavior. Ah, okay. And that data set itself sounds like a major undertaking. Psych 101, you called it. Mm-hmm. Psych 101. And it's just unprecedented in scale. We're talking trial-by-trial data, over 60 ,000 participants, making more than 10 million choices across 160 different experiments. Wow. And the really clever part, I think, was how they standardized it. each experiment, no matter how different, was transcribed into natural language. So you have this common format for wildly different kinds of tasks. So what kind of tasks are we talking about?

2:29What's covered in those 160 experiments? Is it all decision-making? Oh, it's much broader. It covers a whole spectrum. Yeah. Yeah, there's decision-making, like choosing between slot machines, multi-armed bandits, they call them, or selecting lotteries, but also memory tasks, like remembering letter sequences, supervised learning, like predicting weather from tarot cards, believe it or not, even navigating spaceships to planets to find treasure, which are types of Markov decision processes. So all these different human choices, from really simple to quite complex, got fed into Centaur. Okay, so you feeded all this data.

3:01The big question then is, how well does it actually perform? Does it really, you know, capture human behavior? Right. They put it through some really tough tests. First, they looked at held out participants. So these are people whose data Centaur had never seen before, even if it had seen data from the same experiment but from other people. Okay, a standard validation check. Exactly. And Centaur didn't just improve things slightly compared to the base llama model. It significantly improved what they call goodness of fit. Its predictions were just much, much closer to what humans actually chose.

3:33And importantly, it outperformed almost all the existing domain-specific cognitive models in their own experiments. So better than the models designed just for that one test. In nearly every case, yeah. The average difference in log likelihoods, which is a statistical measure of fit, showed a really clear advantage for Centaur. It's pretty remarkable. That's definitely impressive for prediction. But here's where I think it gets really interesting. Can it actually generate human-like behavior? Not just predict the next choice, but simulate a whole sequence of actions, like where its own output feeds back in.

4:07Absolutely. And you're right, that's a much stronger, much harder test. They use something called the horizon task paradigm. It's often used to see if people or models use smart exploration strategies. And Centaur's performance was, well, basically identical to human participants. It scored about 54 points on average. Humans scored around 53. Very close. Very close. And crucially, it showed signs of uncertainty-guided directed exploration. That means when it was unsure, it actively sought out information to reduce that uncertainty. That's a sophisticated human strategy that many other LLMs just don't seem to capture.

4:46And what about the nuances, the different ways humans learn? We don't all approach problems the same way. Exactly right. They looked at the two-step task. This is a classic one, designed to pull apart different types of reinforcement learning. You've got model-free learning, which is more like habit-based stimulus response, and model-based learning, where you build an internal model of the world and plan ahead. Sendor didn't just pick one strategy. It actually replicated the full distribution of human approaches. It generated behavior that looked like pure, pure model-free, pure model-based, and importantly, mixed strategies, which is what you see in real populations.

5:20So it captures the variety within human behavior too. Precisely. But maybe the ultimate test sometimes is, can it fail like a human? Does it make the same kinds of maybe irrational choices or have the same biases? Yes. And this was a really cool finding, I thought. They ran a study where people had to predict choices in economic games. Sometimes they were predicting what another human would do, sometimes what an artificial agent like a bot would do. Okay. Centaur mirrored human performance perfectly. It was good at predicting other humans about 64 % accuracy, same as the real participants. But it struggled to predict the artificial agents only getting about 35 % right, again, just like the humans.

6:00So it's biased towards thinking like a human, even when that's not the optimal strategy for the task. Exactly. It confirms it's not just some perfect predictor. It really does exhibit these human-like characteristics, including our limitations when dealing with things that don't think like us. Okay, this is really compelling. But a true foundation model, you'd think, needs to generalize, right? It needs to handle experiments it's never, ever seen before, completely new situations. How did Centaur do there? Yeah, this is critical. And this is where it really started to shine, I think. They tested it on increasingly difficult out-of-distribution scenarios.

6:36Things truly different from Psych 101. Like what? Okay, first they tried just changing the cover story. So Psych 101 had that spaceship version of the two-step task we mentioned. They tested Centaur on a version with the exact same underlying rules, but framed as writing a magic carpet. Totally different narrative. Right. Centaur still captured human behavior beautifully. It outperformed the base llama model and the traditional cognitive models, even with just that surface-level change. So it wasn't just memorizing the spaceship story. It understood the underlying task structure. Seems like it.

7:08Then they went further. They actually modified the task structure itself. Psych 101 had lots of two-armed bandit tasks, like choosing between two slot machines. They tested it on something called Maggie's Farm, which is a three-armed bandit. A small change, but fundamental. Okay, adding complexity. Yeah. And Centaur, again, robustly captured how humans behaved on Maggie's farm. Meanwhile, the standard cognitive model, the one designed for two-armed bandits, well, it just couldn't generalize. It wasn't built for three arms. That makes sense. It was too specific. But Centaur adapted. It did. But the really big test, the one that addresses, you know, true domain general intelligence, was testing it on something completely outside the scope of Psych 101.

7:52Okay, what was that? Logical reasoning. They used problems similar to what you'd find on the LSAT, the law school admission test. And this is key. The researchers had purposefully excluded any studies involving logical reasoning from the Psych 101 training data. So Centaur had literally zero direct training on this kind of thinking. Exactly. And yet, even on these logical reasoning problems, Centaur showed a significant positive effect from the fine-tuning process. It performed better than the base Lama model. It demonstrated strong generalization to a cognitive domain it had never explicitly encountered during its behavioral training.

8:27That is genuinely mind-boggling. It somehow learned principles of reasoning just by modeling choices and other kinds of tasks. It certainly suggests something deep was learned. And it wasn't just these three examples either. They tested it on six other distinct out-of-distribution paradigms. Things like moral decision-making, different economic games, naturalistic learning tasks. And Centaur consistently captured human behavior across the board, whereas smaller models or the base Lama model, without the fine-tuning, they just didn't perform reliably on these new tasks. Wow. Okay. And you mentioned earlier it wasn't just about the choices people make.

9:01It could also predict how long it took them. It could. And this adds another layer, right? Because response time often reflects cognitive effort, uncertainty, the process itself. They analyzed almost 4 million human response times from the data set. And Centaur's internal representations, basically what's going on inside the model, could explain a significantly larger chunk of the variance in those human response times. More than the base model. More than the base Lanner model, yes. It had a conditional R squared of 0.87 compared to Lama's 0.75. And interestingly, also better than conventional cognitive models specifically designed to predict response times, which scored around 0.77.

9:41So it's capturing something about the process of thinking, not just the final output. That's what it strongly suggests, yes. It's predicting behavior and the timing, which is pretty holistic. Which leads to the next big question. If it's predicting behavior and timing so well, what about its internal workings? Does anything inside Centaur actually resemble how our brains work? Do its representations align with neural activity? Okay, this for me was one of the most exciting and, frankly, unexpected findings in the paper. Remember, Centaur was only trained to match behavior. Choices. Right. Response times, maybe.

10:15Nobody told it to mimic brain activity. Right. It wasn't trained on fMRI data or anything like that. Exactly. But despite that, its internal representation spontaneously became more aligned with actual human neural activity. They did analyses where they tried to predict fMRI measurements. Yeah. Brain scan data showing active regions using the internal states of the models. And Centaur was better at that. Yes. Take the two-step task again. They found that centaur's representations consistently outperformed the base llama model's representations in predicting that human neural activity. This was true across different layers of the model and different brain regions involved in the task.

10:51And this alignment just emerged. It just emerged from the process of fine-tuning on behavior. It wasn't explicitly optimized for. And they saw similar effects, though maybe a bit smaller, even in a totally unrelated task, just reading sentences. It suggests this neural alignment isn't just specific to the decision tasks it was trained heavily on, but might generalize to other cognitive settings too. It's like it stumbled upon brain-like representations, because those are simply efficient ways to produce human-like behavior. That's certainly a compelling interpretation. It suggests there might be some fundamental convergence between how this model learned to solve these problems and how our brains do it.

11:32This is moving beyond just prediction then. It sounds like Centaur, combined with Psych 101, could actually become a tool for scientific discovery itself, like helping us find new theories. Exactly. They actually present a really neat case study showing how this might work. It was in the area of multi-attribute decision making. So, you know, choosing between options based on several features, like picking a phone based on price, camera, battery life. Right. Standard stuff. Right. Researchers usually propose cognitive models for how people do this based on theory or observation. But here, they use AI to help discover a new model.

12:07How did that work? This sounds really interesting. Well, first, they used another large language model called DeepSeek R1. Its job was to analyze the human data from Site 101 for a specific multi-attribute task and just generate verbal explanations like hypotheses for how people might be making those decisions. So using one AI to generate initial ideas. Precisely. And DeepSeek R1 actually came up with a novel strategy. It proposed a two-step heuristic, basically combining two known cognitive shortcuts, two simpler rules of thumb, in a way that hadn't really been formally considered together before.

12:42And this new strategy, proposed by DeepSeq R1, actually did a better job explaining the human data than the traditional models used in the original study. Okay, so AI already helped find a better interpretable model, but then Centaur came back into the picture. Right. Because as good as the DeepSeq R1 model was, Centaur's predictions were still slightly more accurate overall. It was capturing some nuance the heuristic model missed. Ah, so Centaur becomes the benchmark. Exactly. So they used this method they called scientific regret minimization. Fancy name. But the idea is simple. They looked specifically at the choices where Centaur was right, but the Deep Seek R1 heuristic model was wrong.

13:19Where did they disagree? Okay, looking for the errors of the simpler model. Yeah. And by inspecting those specific cases, they found a subtle pattern. It seemed like participants weren't always strictly switching between the two heuristics as DeepSeek R1 proposed. Sometimes they'd pick an option that maybe had fewer good ratings overall. If one of its positive ratings came from a source they considered more reliable, a higher validity expert, basically. Ah, so the switch wasn't strictly either or. It was more graded, more flexible. Exactly. It suggested people weren't just using heuristic A or heuristic B, but maybe blending them.

13:53So what did that mean for refining the DeepSeq R1 model? They took that insight and modified the model. Instead of a strict switch, they implemented a weighted average of the two heuristics. The resulting model, this refined heuristic model, it not only matched Centaur's excellent goodness of fit, its accuracy, but it remained simple and interpretable. It gave a new scientific insight. People likely use this dynamic combination of heuristics, not just one strategy or the other. Wow. So you have AI generating a hypothesis, a more powerful AI refining it by spotting discrepancies and humans interpreting the result into a new cognitive theory.

14:31That's the blueprint they demonstrated. It's a powerful way to potentially accelerate scientific discovery using these models as tools. OK, this deep dive has really shown Centaur as well, an incredibly powerful new tool. Thinking bigger picture now, what does this all mean for the future of, you know, understanding ourselves? Well, there are some really direct opportunities that spring from this. The paper talks about enabling automated cognitive science. One aspect is what they call in silico-trototyping. Imagine using Centaur to simulate experiments before you run them on actual people. Like testing out different designs.

15:03Exactly. You could use Centaur to predict which experimental design might give you the biggest effect size, or figure out how many participants you might need to get enough statistical power, or even just estimate if an effect is likely to exist before you invest time and resources recruiting people. That would be hugely practical for researchers. Save a lot of time, a lot of money. Absolutely. But beyond the practical, there are deeper theoretical questions this opens up. Such as? Well, now that we have this model whose internal states seem to align somewhat with neural activity, we can start to probe those representations directly.

15:36We could use techniques like sparse autoencoders or attention map visualization ways to sort of look inside the model to generate new concrete hypotheses about how humans represent knowledge or process information. And then test those hypotheses on real people. Precisely. It creates this loop between computational modeling and experimental work. We could even try training different kinds of models, maybe with different architectures, from scratch on Psych 101 to see which types of computational structures best capture human information processing. And the data set itself, Psych 101, that's not static either, is it?

16:13It's growing. That's right. It's an ongoing project. The current version is heavily focused on learning and decision making, which are foundational. But the plan is to expand it, bring in data from psycholinguistics, social psychology, more economic games. Drawning the scope of cognition covered? And also incorporating individual differences, things like age, personality traits, socioeconomic status, and critically, addressing the current bias in a lot of psychological research towards weird populations. Weird being Western, educated, industrialized, rich, and democratic. Exactly. So making the data set, and thus the models trained on it, more representative of global human diversity.

16:53They even mentioned the long-term goal of a multimodal data format, perhaps incorporating things beyond just choices. So it feels like we're really just at the beginning of what this kind of approach could achieve. I think so. Okay, let's try and wrap this up. We've dived deep into Centaur today, a truly groundbreaking foundation model for human cognition. We saw how it was fine-tuned on this massive, diverse Psych 101 data set, how it learned not just to predict human choices with incredible accuracy, but also to simulate human-like behavior, including exploration and different learning styles.

17:27And how it generalized remarkably well to completely new tasks, even logical reasoning it wasn't trained on. Plus that amazing finding about its internal representations spontaneously aligning with human neural activity. Yeah, that was really something. And then how it can even be used as a tool for scientific discovery, helping refine our cognitive theories. It really feels like that old vision of a unified theory of cognition, which maybe seemed abstract or even like an intruder, into specialized fields. Well, it's starting to become much more concrete through these kinds of large-scale data-driven models.

18:00Centaur basically aced the equivalent of, what, 16 different cognitive decathlons consistently beating the specialized champions. So the big question left hanging then is how we translate this incredibly powerful computational model into a truly satisfying unified theory of the human mind. And maybe for you listening, what does it mean for your understanding of yourself, of others, knowing that we're beginning to map the very processes of human thought in this computational way. Something to think about. Thank you for joining us on this deep dive. We hope you feel a little more well-informed and definitely a lot more curious.

18:33Until next time.