A small cognitive core explains human behavior at scale

Introduction

• The open question is how much human behavior established cognitive principles explain together, despite extensive knowledge of individual mechanisms.A proposed test is a cognitively meaningful model evaluated across a broad range of experiments.Source: Introduction · p. 2
• CORE combines a hierarchy of episodic memories, prediction-error writing, and context-dependent encoding, retrieval, and forgetting.It operates on natural-language transcripts of experimental sessions, enabling fitting across paradigms.Source: Introduction · p. 2
• Across 137 experiments and 7,036,442 choices, CORE predicts choices well against Transformers and Centaur, and simulations reproduce 75% (50 out of 67) of reported behavioral effects.The model was fitted jointly to the experiments; the effects were reproduced in simulation.Source: Introduction · p. 3
• The authors argue that a small cognitive core may explain much of human behavior and provide a framework for testing theories against large portions of the behavioral record.Source: Introduction · p. 3

Results

• CORE processes experiment transcripts as phonemes, then uses stacked episodic-memory layers to encode context-dependent associations and retrieve values for predicting button presses.Each layer computes keys, queries, forgetting, and learning from its current input; prediction errors, rather than raw inputs, are written to memory.Source: Results · pp. 4–5
• CORE achieved lower NLML (3,498,133 vs. 4,592,031) and PNLL (262,572 vs. 329,043) than the Transformer, and lower PNLL (262,572 vs. 389,151) than Centaur.It predicted held-out choices better than the Transformer in all 62 studies and better than Centaur in 47.Source: Results · p. 6, Figure 2 · p. 7
• Ablations showed that hierarchy mattered most (ΔNLML = 859,990; ΔPNLL = 98,916), followed by forgetting, phoneme-based encoding, and writing prediction errors rather than raw inputs.Without hierarchy, the model reproduced 22 of 67 effects; without forgetting, it reproduced 42.Source: Results · pp. 8, 10
• 50 of 67 behavioral effects (75%) were reproduced in simulations, with the strongest match in exploration and reinforcement learning and the weakest in language and social judgment.Six missed effects were in the reported direction but nonsignificant; some were also fragile in human resamples.Source: Results · pp. 8, 10, Table 1 · p. 9
• Learning rates fell with layer depth (median 0.44 in the first versus 0.17 in the last layer), while median memory timescales rose from 8 tokens in layer one to 43 in layer four.Lower layers were associated with recent input and higher layers with broader, more context-specific processing.Source: Figure 3 · p. 11, Results · p. 12
• Learning rates declined in 89 of 128 dimensions, while memory timescales grew within games and shortened at game boundaries in 94 of 128 dimensions.The strongest plotted timescale change was an 80% drop at a game boundary; the effect occurred in all 17 experiments.Source: Figure 3 · p. 11, Results · pp. 12–13

• Psychology’s existing principles often fail to generalize beyond the paradigms they were developed for, motivating tests of their combined explanatory reach.Source: Discussion · p. 13
• CORE predicts human choices better than standard machine-learning architectures and a foundation model of human cognition with five orders of magnitude more parameters, and reproduces 75% of behavioral effects reported in original studies.Source: Discussion · p. 13
• CORE brings together established phonemic input, associative memory, context dependence, prediction errors, and hierarchical representations, with proposed neural implementations for associative memory [46,47].Source: Discussion · pp. 13–14
• CORE contains no computation beyond memory writing and retrieval, making its behavior attributable to a handful of principles, though its computations are not interpretable.Source: Discussion · pp. 14–15
• CORE struggles with model-based planning, and it has no explicit planning mechanism.The authors identify adding a planning mechanism as one possible direction for future model iterations.Source: Discussion · p. 14
• Experiments typically last less than an hour, so the available data do not reveal how participants learn and store information over weeks or months.The authors identify longitudinal data as a way to study long-term memory and its relation to episodic memory.Source: Discussion · p. 14
• CORE operates directly on natural language and could be extended to other sensory modalities, while traditional cognitive models resist naturalistic environments.Source: Discussion · pp. 14–15

• CORE shows that a small set of established cognitive mechanisms can account for a broad range of behavior when fitted jointly across experiments.Source: Conclusion · p. 15
• The authors propose that a shared model constrained by every experiment could make psychology more than the sum of its experiments.Source: Conclusion · p. 15

• The re-transcribed Psych-201 data comprise 137 experiments from 62 studies, with 7,634,315 choices in 83,352 sessions; 597,873 choices in 2,900 sessions were held out for testing.Transcripts were produced with an agentic pipeline, and sessions without button presses were dropped.Source: Materials and Methods · p. 16
• CORE maps phonemes to learned 128-dimensional embeddings processed by six memory layers, each with four parallel heads; the model has 452,608 parameters.Each head computes keys, queries, learning rates, and forgetting factors; a linear readout and softmax predict response buttons.Source: Materials and Methods · pp. 16–17
• Parameters were fitted jointly for 100 epochs using maximum a posteriori estimation, with training taking about 24 hours on one NVIDIA H100 GPU.Source: Materials and Methods · p. 17
• Predictive log-likelihood on held-out choices was compared with a matched Transformer and component ablations; Centaur’s marginal likelihood was not computed because accounting for its pretraining was unclear.Ablations removed memory depth, forgetting, error-driven encoding, or phoneme tokenization.Source: Materials and Methods · p. 17
• Simulations tested 67 effects from 47 experiments across 32 studies, retaining effects reproducible in the original human data and analyzable from simulator outputs.CORE’s predicted distributions were sampled at temperature 1.0, with up to 1,000 simulated participants per study.Source: Materials and Methods · p. 18
• Latent analyses recorded learning rates and forgetting factors; forgetting was converted to memory timescales, defined as tokens until an association decayed to 1/e of its strength.Source: Materials and Methods · p. 18
• For latent-variable distributions, learning rates and log forgetting factors were averaged across each head’s dimensions, then histograms were normalized and averaged equally across experiments.Source: Materials and Methods · p. 18
• For game-sequence analyses, choices were grouped into five relative-position bins and averaged within participants, then across participants and experiments with equal experiment weighting.Games with fewer than five choices were excluded; error bars show standard error across experiments in within-participant change.Source: Materials and Methods · pp. 18–19