A fine-tuned Llama on psych trials to predict human choices. Literally beats specialist models, generalizes to new tasks, and even echoes fMRI patterns.
Imagine you want to understand how people make decisions. Not just in one specific scenario, like choosing a coffee brand, but across a huge range of situations.
For decades, psychologists (and behavioural economists) have built specialized models for each of these scenarios or "domains."
A model for risk
Another model for learning, or vicarious learning (Bandura's Social Cognitive Theory)
A model for memory…
Each crafted for one single domain.
But what if you could build a swiss army knife kind of model? A single, general model that captures the underlying patterns of human choice, no matter the context.
A couple of weeks ago I came across a new paper, "A foundation model to predict and capture human cognition", introduces exactly that. They call this swiss army knife model Centaur.
At its heart, Centaur is a language model (Meta's Llama 3.1 70B) that has been taught to think a little more like a person. More specifically, how the make choices.
Centaur starts as Meta’s Llama 3.1-70B. Researchers then add tiny low-rank adapters (rank-8, ~0.15% extra params) and fine-tune on a massive, hand-transcribed corpus of classic psych experiments, 160 studies, 60,092 participants, 10,681,650 choices, ~254M tokens, where each trial is written as a little story (“You see … You press <<B>>”). One epoch, standard cross-entropy, and the loss is only applied to the tokens that encode the human’s choice. Training took ~5 days on an A100 80GB. (Vast.ai A100 PCIe: $0.67–$1.40/hr → $80–$168 for 5 days)
You can't teach an AI about human cognition from Wikipedia, or Reddit. You need to show it what people actually do. The researchers transcribed a massive dataset called Psych-101. They took the raw data from 160 different psychological experiments (over 10 million individual choices from 60,000 people) and translated each one into a simple story. The AI doesn't just see a spreadsheet of 0s and 1s. It reads a running log of the entire experiment, just as a human participant would experience it.
Here’s a simplified look at what one of these "stories" from the dataset looks like:
# Instructions
In this task, you have to repeatedly choose between two slot machines labelled B and C. When you select one of the machines, you will win or lose points. Your goal is to choose the slot machines that will give you the most points.
# Trial History
You press <<C>> and get -8 points.
You press <<B>> and get 0 points.
You press <<B>> and get 1 points.
You press <<C>> and get 5 points.
You press <<C>> and get 2 points.
You press <<B>> and get -1 points.
...and so on...The model reads this entire history and its task is to predict the next choice, like <<B>>.
This is where it gets really interesting. The researchers put Centaur through a series of tests to see if it had actually learned something fundamental about cognition.
It Beats the Specialists. They pitted Centaur against the classic, hand-crafted cognitive models on their own domains. In nearly every case, Centaur was better at predicting the choices of held-out participants.
This is a big deal…
The general purpose swiss army knife was outperforming the specialized screwdrivers. It learned better representations of human behavior from the data alone than the models explicitly designed for that task.
It Generalizes. This was the acid test. Can it handle something it's never seen before?
They gave it the same task but with a different theme (e.g., "magical carpets" instead of "spaceships")??. Centaur's performance held up. It understood the underlying task structure, not just the surface-level story.
They tested it on a task with three choices instead of two. Even though it was mostly trained on two-choice tasks, it successfully adapted.
They even tested it on tasks like logical reasoning, which were intentionally excluded from the training data. It still worked. This suggests it's not just memorizing, but capturing a more abstract, flexible model of cognition.
This is the wildest part. Its internals drift toward brain-like patterns.
They compared fMRI scans of humans doing a task to Centaur’s internal activations while “thinking” about the same task. After fine-tuning on behavior, the alignment increased. Training only on choices nudged the model’s representation geometry closer to real neural activity.
Let that sink in for a bit…
By only training the model to predict a person's choices, its internal "thought process" started to spontaneously resemble the patterns of a real brain. This might suggest that our patterns of choice are a deep reflection of our neural architecture, and Centaur was able to pick up on that echo.
The small sibling
There’s a smaller variant Minitaur-8b that you can run locally. You can try it and run it yourself using LM Studio
Minitaur running locally on my machine. Minitaur has correctly identified that Machine B is paying out more (the "exploit" choice). It has learned the fundamental "explore-exploit" dynamic from its training and will act accordingly.
Because Minitaur hasn't been explicitly trained on logical deduction, its performance might be less reliable. However, its prediction is correct <<Valid>>.
You will be shown a series of marketing emails sent to a user.
Your task is to learn the user's rule for opening an email.
For each email, you will see the subject line and whether the user chose to open it.
Subject: "Welcome to Vanguard! Here's 15% Off Your First Order"
Your action: <<Open>>
Subject: "New Arrivals: The Summer Collection Is Here"
Your action: <<Ignore>>
Subject: "FLASH SALE: 40% Off All Jackets, Today Only!"
Your action: <<Open>>
Subject: "A Letter From Our Founder"
Your action: <<Ignore>>
Subject: "Your Last Chance for 25% Off Sitewide"
Your action: <<Open>>
Subject: "Style Guide: How to Wear Linen This Season"
Your action: <<Ignore>>
Subject: "It's a Sale on a Sale! Extra 20% Off Clearance"
Your action: <<Open>>
Subject: "20% Off Is Back!"
Your action: <<Open>>
Subject: "Spotted: Celebrities Wearing Vanguard Apparel"
Your action: <<Open>>
Subject: "Clearance Update: New Markdowns Added"
Your action: <<Ignore>>
Subject: "Vanguard Insiders Get Early Access..."
Your action: <<Open>>
Subject: "Another Sale? You Bet. 15% Off Everything."
Your action: <<Ignore>>
Subject: "Behind the Scenes of Our Fall Photoshoot"
Your action: <<Open>>
Subject: "Your Exclusive Invite: The Vanguard VIP Program"
Your action: <<Open>>
You receive two new emails at the same time.
- Email A Subject: "It's Friends & Family! Take 30% Off"
- Email B Subject: "You're Invited: First Look at the Limited Edition Capsule"
What is the user's next action? Which email will they open? <<Email A>> or <<Email B>>?Minitaur recognized that the user has "graduated" from bargain hunting and is now motivated by feeling special and having access that others don't. The perceived value of exclusivity has surpassed the value of a discount for this specific user at this point in their lifecycle.
Okay, so we have a magic black box that's great at predicting people’s choices.
But, what can you actually do with it?
They did a case study where they wanted to understand how people make choices when faced with multiple attributes (e.g., choosing a product with 4 different expert ratings).
They first prompted another AI (Deepseek-R1) to look at the human data and describe the strategy in plain English. It came up with a reasonable two-step rule.
They formalized that rule into a simple model. It was okay, but Centaur was much better at predicting what people would do. They then looked at the specific choices where the simple model failed but Centaur succeeded.
By analyzing these specific cases, they figured out what was missing from the simple rule. People weren't using a strict either/or rule, but a more fluid combination of strategies. They built this new insight into a new, interpretable model.
This resulted in a brand-new cognitive model, discovered with the help of an AI, that was just as predictive as the Centaur black box but is also simple enough for a human scientist to understand.
They used the LLM not as the answer, but as a sort of "microscope" to find the subtle patterns they had missed, leading to better, more accurate theories about ourselves.
I believe this is another glimpse into the future, "in silico" science (see also Alpha Fold and Alpha Genome), where we can prototype experiments, generate hypotheses, and refine theories about human behavior at a scale and speed that was previously unimaginable.