A memory hierarchy that writes prediction errors
CORE, short for COntextual Retrieval and Encoding, is not a new principle. It is a specific assembly of old ones: an episodic memory that stores experiences, retrieval driven by context, learning driven by prediction errors, forgetting, and a hierarchy of representations. The model reads the natural-language transcript of an experimental session — instructions, stimuli, feedback, and the participant's own earlier responses — and outputs a probability for each possible button press.
The concrete mechanism is an associative matrix. At each step the model computes a key, a query, a forgetting factor, and a learning rate from the current input, so what gets stored, retrieved, and decayed all depend on context. What is written is not the raw input but the prediction error: the difference between the input and what memory already predicts for that key. Several such memory layers are stacked, and a linear read-out followed by a softmax turns the final layer into button-press probabilities.
Two design choices make the cross-paradigm fit possible. Transcripts are converted into phonemes rather than the byte-pair tokens standard in language models, and the model contains no computation beyond writing to and reading from memory. That second point matters for interpretation: because nothing else is in the architecture, its behavior can be attributed to those principles and little else.
Joint fitting across 137 experiments
The training data come from Psych-201, a large collection of psychological experiments spanning reinforcement learning, decision making, category learning, and social interaction. The analyses focus on the 137 experiments from 62 studies in which participants responded by button press, comprising 7,036,442 choices across 80,452 experimental sessions. CORE was fitted to all of them jointly using maximum a posteriori estimation.
Two metrics carry the comparison, and they answer different questions. Negative log marginal likelihood (NLML) measures how improbable the data are when averaged over all possible model weights weighted by their prior, so it reflects model evidence rather than pure prediction. Predictive negative log-likelihood (PNLL) measures how well the fitted model predicts human choices in held-out sessions. Lower is better on both, and the paper reports them separately rather than collapsing them into one score.
The baselines are a Transformer-based sequence model of matched configuration, Centaur — a large language model fine-tuned on human behavior that the paper describes as outperforming domain-specific cognitive models — and ablated versions of CORE with one component removed at a time.
Better predictions than a Transformer and Centaur
CORE achieved an NLML of 3,498,133 against 4,592,031 for the matched Transformer, and a PNLL of 262,572 against 329,043. Against Centaur the held-out comparison was 262,572 versus 389,151; Centaur has no NLML entry in the figure because it is not fitted the same way. The paper notes that Centaur has five orders of magnitude more parameters than CORE.
The aggregate numbers are backed by per-study counts. CORE predicted held-out choices better than the Transformer in all 62 studies, and better than Centaur in 47 of the 62. So the advantage over the Transformer is uniform across the included studies, while the advantage over Centaur is broad but not universal — in 15 studies Centaur was at least as good on this metric.
Removing any component costs predictive accuracy
Ablations show which pieces carry the result. Removing the hierarchy was by far the largest loss (change in NLML of 859,990 and in PNLL of 98,916); with a single layer, the paper states, the model lacks the expressivity to capture human behavior. Removing forgetting came second (131,706 and 13,483), because without it old associations never fade and can only be overwritten.
The encoding choice also mattered. Phoneme-based encoding beat byte-pair encoding (110,303 and 7,014), even though the byte-pair model fit the training data better — a training negative log-likelihood of 2,904,051 versus 3,066,566 — and used far more embedding parameters, 1,750,016 versus 11,136. Better training fit did not translate into better generalization here.
Finally, writing prediction errors instead of raw inputs improved performance (68,874 and 4,423). The paper's explanation is capacity: raw inputs re-add familiar content, whereas prediction errors carry only what memory has not yet learned.
Reproducing 75% of effects and classic learning signatures
Prediction accuracy alone does not show that a model behaves like a participant. When simulated, CORE reproduced 75% of the behavioral effects identified in the corresponding studies — 50 out of 67. The remaining quarter is where the paper locates its main gap.
The latent variables were also inspected during simulated experiments. Learning rates decreased with depth (median 0.44 in the first layer versus 0.17 in the last), and memory timescales increased from the first to the fourth layer (median 8 versus 43 tokens). Lower layers thus track recent input while higher layers form slower, more integrative averages; lower layers showed narrow unimodal distributions and higher layers broad, often multimodal ones.
Two classical signatures appeared in experiments made of sequences of independent games. Learning rates declined as evidence accumulated, as Kalman-filter-style Bayesian accounts predict: the decline held in 15 of 17 experiments in the strongest dimension and on average in 89 of the layer's 128 dimensions. Memory timescales grew within a game and collapsed at game boundaries, consistent with event segmentation: an 80% drop at the boundary in the strongest dimension, robust in 17 of 17 experiments, and shorter boundary timescales in 94 of 128 dimensions on average.
No planning, and no memory beyond the session
The paper is explicit that CORE struggles with model-based planning, which follows from containing no explicit planning mechanism; the authors suggest future iterations could add one. This is the most direct candidate for part of the unexplained 25% of behavioral effects.
A second boundary comes from the data rather than the architecture. Experiments in the dataset typically span less than an hour, so learning and storage over weeks or months is never observed, and a long-term memory cannot be identified from these transcripts. The authors note that longitudinal, naturalistic data would be needed to add one and to study consolidation.
There is also an interpretive limit the authors state themselves: CORE's principles are interpretable but its computations are not. How it computes keys, values, learning rates, and forgetting lacks theoretical grounding, so the model is best read as an intermediate step between classical cognitive models and black-box systems rather than a full mechanistic account.
A breadth test, not a claim of completeness
The paper's distinctive move is not a new mechanism but a test of combined explanatory power. Prior work offered domain-specific cognitive models that are interpretable but confined to one paradigm, and large-scale predictive models that generalize across paradigms but stay black-box. CORE is judged against both, and it wins on held-out choices against both while remaining built from named principles.
That makes the per-study split the most decision-relevant number for a reader weighing the claim. Uniform superiority over the matched Transformer across 62 of 62 studies supports the architecture comparison; superiority over Centaur in 47 of 62 supports the claim that a small principle-based model can compete with a vastly larger black-box one, but not that it dominates it everywhere.
The honest summary is therefore narrower than "a small cognitive core explains human behavior." It explains a large share of button-press choices across 137 experiments, reproduces three quarters of the reported effects, and shows learning-rate and event-boundary signatures the authors interpret as Bayesian learning and event segmentation. Planning and timescales beyond a single session remain outside what this evidence covers.
Sources used in this answer
A small cognitive core explains human behavior at scale
CORE, a hierarchy of context-dependent episodic memories that store prediction errors, was fitted jointly to 7,036,442 choices from 137 experiments and predicted held-out human choices better than a matched Transformer and Centaur, while reproducing 75% (50 of 67) of reported behavioral effects in simulation.