A small cognitive core explains human behavior at scale

arXiv:2610.13603 · cs.LG · Submitted 2026-10-09 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "A small cognitive core explains human behavior at scale".

Jane: The gist The COntextual Retrieval and Encoding (CORE) model, which combines established cognitive principles into a single framework,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So, Jane, this paper, "A small cognitive core explains human behavior at scale," it's about trying to build one unified model instead of just testing principles in isolation. It claims they assembled a small set of established cognitive principles into something called CORE which is Contextual Retrieval and Encoding.

Jane: Right, so the big idea here is that instead of having many separate theories about how we think, they put them together into one system that tries to explain human behavior across different kinds of experiments. It’s fitting this single model against a huge amount of data to see if it actually works better than existing systems.

Tom: Exactly, and what they found is that this CORE model predicts human choices better than some standard machine learning architectures like Transformers and even black-box foundation models like Centaur. They also say when they simulate this CORE model, it reproduces seventy-five percent of the reported behavioral effects from the experiments they tested <ref:2610.13603#pg1>.

Lu: That’s a pretty strong claim because it suggests that this small set of principles actually brings us quite far in explaining human behavior according to their findings, and what's really interesting is that their latent variables show signatures of Bayesian learning and event segmentation.

Meng: If you're thinking about deploying something like this, the focus would be on how much compute it needs to run for all these one hundred thirty-seven experiments, because fitting it jointly to seven million choices across eighty thousand sessions sounds like a lot of work <ref:2610.13603#pg3>.

Lalam: I see what they mean by that small set of principles being key; they found that the hierarchy in CORE is actually the most important component, with the ablation study showing a big jump in performance when keeping that structure.

Tom: That’s interesting, so even if you take away the forgetting mechanism, which is supposed to keep things contextual, the model still only gets forty-two of those sixty-seven effects reproduced. It shows how much context dependence matters for this system to work at scale.

Jane: So what does this mean for us on a normal day? It means that our brains might be using these same basic cognitive mechanisms—like remembering things and making predictions based on context—in a way that is more organized than we thought, even though the model itself doesn't have any explicit planning mechanism yet.

Lu: That’s one of their main caveats, though they admit the model struggles with model-based planning because it doesn't have an explicit planning mechanism built into it right now. They also mention that longitudinal data would be needed to study long-term memory, which isn't observed in the short experimental spans they used.

Meng: For a production system, the lack of an explicit planning mechanism means you can’t just feed it a goal and expect it to figure out the steps sequentially; you’d need to build that planning layer on top of this cognitive core later.

Lalam: The paper suggests that the opportunity here is larger than any single model because if every experiment constrains a shared model of the mind, then psychology might finally become more than just a collection of experiments.

Tom: So we're moving toward assembling these established cognitive principles into one big model that gets fitted jointly to many experiments instead of testing them one by one, and that’s what this work is about with "A small cognitive core explains human behavior at scale."

Jane: It really suggests a direction where we can move away from just testing theories in isolation toward having an actually unified account of the human mind.

Lu: It's a step between classical cognitive models and black-box systems, which is useful because it shows that these established cognitive mechanisms can account for surprisingly broad human behavior when they are fitted jointly across experiments.

Meng: For what we deploy today, it’s an intermediate step; it’s not the final product because it lacks the explicit planning capability needed for complex sequential tasks.

Lalam: So if you're tuning in just to get the main idea, think of CORE as proving that a small set of established cognitive mechanisms can explain a surprisingly broad range of human behavior when they are fitted jointly across experiments.

Conclusion: Tom: So, we're wrapping up our talk on "A small cognitive core explains human behavior at scale." It turns out this paper is about taking a bunch of established cognitive ideas and putting them into one single framework to try and predict human choices better than these big black-box models.

Jane: Right, so the authors are showing how they fit a small set of rules—like memory layers and how information gets forgotten—onto massive amounts of experimental data. It's trying to see if this simple cognitive structure can actually reproduce real human behavior across many different kinds of tasks.

Lu: The core idea is that these established cognitive mechanisms, when put together right, can explain a surprisingly wide range of human actions when you fit the model jointly to experiments from all these different paradigms.

Meng: But they're admitting there's still a gap; this core system doesn't have any explicit planning mechanism built in yet, so it can’t do complex, multi-step reasoning on its own right now.

Lalam: That’s the main point; it shows that if you constrain a single model of the mind with many experiments, then psychology might finally become more than just a collection of separate studies.

Tom: It really is about assembly rather than isolation here. This work suggests that we can move toward building one unified account instead of just testing theories in isolation.

Jane: And what this means for us is that we're moving toward systems where basic cognitive processes are seen as a whole, and that's a direction worth paying attention to.

Lu: It’s an intermediate step, though; it’s not the final answer because they haven't added explicit planning yet. Future work will need to tackle how this core handles long-term memory across longer time spans.

Meng: The focus now has to be on making sure this core is robust enough for real-world deployment, because right now, it’s more of a research tool than something you can just plug into a production system.

Lalam: It's a step between classical cognitive models and these black-box systems we see everywhere. We need to see if this cognitive foundation is strong enough to handle the complexity of real human decision-making.

Marcel Binz, Eric Schulz

Institute for Human-Centered AI, Helmholtz Munich · Max Planck Institute for Biological Cybernetics

cs.LG

Submitted: 2026-10-09

Updated: 2026-10-09

Code: https://github.com/fla-org/flash-linearattention

Project page: https://core-cognition.github.io/core-model/results

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 89/100

The gist: The gist The COntextual Retrieval and Encoding (CORE) model, which combines established cognitive principles into a single framework, predicts human choices better than black-box models and

Key concepts

CORE Model: A small cognitive core that uses a hierarchy of episodic memories to predict human choices. It processes experimental transcripts by converting them into phonemes and embedding them through memory layers where context influences how information is stored, retrieved, and forgotten.
Context-Dependent Mechanisms: The model's encoding, retrieval, and forgetting processes all depend on the current context. This means what information is stored or remembered changes based on the specific situation or 'key' being processed at any given time.
Hierarchical Representations: The model uses multiple layers of abstraction in its memory structure. Lower layers track recent inputs, while higher layers form running averages that capture the overall structure of an experiment. This hierarchy allows it to recognize both specific tokens and broader sequential contingencies.
Prediction Errors: Encoding involves writing 'prediction errors' into the memory. This error is defined as the difference between what the model actually received in a given input and what its memory previously predicted for that specific context, driving learning.

Terminology

Core — 1× in this paper · explained in 2 episodes

Summary

The gist The COntextual Retrieval and Encoding (CORE) model, which combines established cognitive principles into a single framework, predicts human choices better than black-box models and reproduces 75% of reported behavioral effects when simulated.

Model Architecture

CORE is a small cognitive core that consists of a hierarchy of episodic memories into which prediction errors are written and whose encoding, retrieval, and forgetting depend on context. The model operates directly on natural language transcripts of experimental sessions, allowing the same model to be fitted across paradigms. The architecture involves converting the transcript into a sequence of phonemes, which are then embedded and passed through a stack of memory layers where each layer adds its output to its input. Each memory layer computes a key, a query, a forgetting factor, and a learning rate from the current input to make every operation contextual.

Core Mechanisms

The encoding process involves writing prediction errors into the memory, where the value is defined as the difference between the input and what the memory predicts for the current key. Retrieval occurs when a query reads out values whose keys match it, weighted by similarity. The memory update follows a formula where the existing memory content Ht−1 is multiplied by a feature-wise forgetting factor αt and a new association between a value vt and a key kt is encoded using a Hebbian outer-product rule. The model relies on context dependence, as what is being stored, retrieved, and forgotten should depend on the current context.

Performance and Validation

Fitted jointly to 7,036,442 choices from 137 experiments in 80,452 experimental sessions, CORE predicts human choices better than standard machine learning architectures like Transformers and black-box foundation models such as Centaur. Furthermore, when simulated, CORE reproduces 75% (50 out of 67) of behavioral effects identified in the corresponding studies. The model captures human choices across paradigms, achieving a lower negative log marginal likelihood and predictive negative log-likelihood than the Transformer in all 62 studies and better than Centaur in 47 of them.

Latent Variable Interpretation

CORE’s latent variables show classic signatures of Bayesian learning and event segmentation, with lower layers showing domain-general processes that are invariant across contexts. Specifically, "learning rates decreased with depth (Fig. 3A; median 0.44 in the first vs. 0.17 in the last layer), indicating that lower layers track the most recent input, whereas higher layers form running averages that capture the structure of an experiment". These latent variables exhibit signatures of classical learning phenomena, as CORE’s learning rates also exhibited a decline predicted by Bayesian models of learning, such as the Kalman filter.

Implications and Future Directions

CORE demonstrates that a small set of established cognitive mechanisms can account for much of human behavior. However, the model struggles with model-based planning because it contains no explicit planning mechanism. Future iterations could incorporate such a mechanism, and longitudinal data would allow for the study of long-term memory, which is currently not observed in the short experimental spans. CORE serves as an intermediate step between classical cognitive models and black-box systems.

Conclusion

CORE shows that a small set of established cognitive mechanisms can account for a surprisingly broad range of human behavior when fitted jointly across experiments. The opportunity is larger than any single model, suggesting that if every experiment constrains a shared model of the mind, psychology may finally become more than the sum of its experiments.

How it works

The CORE model reads a transcript of an experimental session and converts it into a sequence of phonemes in line with models of speech perception. Each phoneme is mapped onto a learned d-dimensional embedding, and these embeddings pass through six memory layers where each layer adds its output to its input. The encoding process involves writing prediction errors into the memory, which is defined as the difference between the input and what the memory predicts for the current key.

Key Components

The model's components include:

A hierarchy of episodic memories into which prediction errors are written.
Context-dependent encoding, retrieval, and forgetting.
Phoneme-based input representation.
Hierarchical representations reflecting multiple levels of abstraction.

Evaluation Metrics

The model is evaluated using two metrics: the negative log marginal likelihood (NLML) and the predictive negative log-likelihood (PNLL). CORE achieves a lower NLML and PNLL compared to standard machine learning architectures like Transformers and foundation models like Centaur. The simulation of CORE on Psych-201 reproduced 50 of the 67 behavioral effects, indicating it mirrors qualitative patterns of human behavior.

Ablation Study

The ablation study revealed that the hierarchy proved to be the most important component (∆ NLML = 859,990, ∆ PNLL = 98,916). Removing the forgetting mechanism resulted in a model that reproduced 42 of the 67 effects. The variant without hierarchy reproduced only 22 of the 67 effects, showing that a single layer can only recognize familiar tokens but not learn sequential contingencies between choices and outcomes.

Conclusion

CORE is an initial step in a direction where principles are no longer tested in isolation, but assembled into a single model that is fitted jointly to many experiments. It shows that a small set of established cognitive mechanisms can account for a surprisingly broad range of human behavior when fitted jointly across experiments.

Materials and Methods

The data used is a version of Psych-201 re-transcribed by an agentic pipeline, resulting in 7,634,315 choices in 83,352 sessions. Each button press was represented by a single token to ensure that a button press is never split into phonemes or subwords. The model contains no other components and has 452,608 parameters in total. Training was performed using maximum a posteriori estimation with schedule-free AdamW. The model comparison approximated the log marginal likelihood using a Laplace approximation around the fitted weights.

Experiment Simulations

For each experiment, an agentic pipeline produced a simulator of the task and an analysis script that tests the main effects reported in the original paper. The simulations retained only effects that the original human data reproduce under the same analysis. In total, this yielded 67 effects from 47 experiments of 32 studies.

Improvements for AI systems

A new model can be built that predicts human choices better than black-box models by combining a small set of established principles into one model, which we call COntextual Retrieval and Encoding (CORE). This system will perform joint fitting across multiple paradigms, moving beyond domain-specific or purely predictive architectures.
The improved system can reproduce 75% of reported behavioral effects when simulated, suggesting it will accurately simulate human decision-making across various experimental settings by modeling the hierarchy of episodic memories and context-dependent encoding, retrieval, and forgetting.
The AI system can be enhanced to capture latent variables that show classic signatures of Bayesian learning and event segmentation, allowing it to infer underlying cognitive processes like learning rates and memory timescales from behavioral data.
The system can be adapted to incorporate longitudinal data by extending CORE with a mechanism for long-term memory, enabling the study of how information is consolidated across weeks and months, which is currently beyond its scope.
The AI can be refined to address limitations in model-based planning by incorporating mechanisms, such as those from looped Transformers, to capture complex sequential reasoning that CORE currently struggles with.

Abstract

Psychology has proposed many principles of cognition. Each, however, was developed in isolation, leaving their combined explanatory power unknown. We assembled a small set of established principles into one model, which we call COntextual Retrieval and Encoding (CORE). CORE consists of a hierarchy of episodic memories into which prediction errors are written and whose encoding, retrieval, and forgetting depend on context. Fitted jointly to 7,036,442 choices from 137 experiments, the model predicts human choices better than black-box models and reproduces 75% of reported behavioral effects when simulated. Moreover, its latent variables show classic signatures of Bayesian learning and event segmentation. Taken together, these findings suggest that a small cognitive core accounts for much of human behavior. They also open a path toward making psychology more than the sum of its parts, by building computational models on the entire experimental record.