Individual differences need not each get a model: a six-dimensional shared structure in 155 click-task participants replicates in 102 more

Individual differences in perceptual decision-making collapse into different weightings over a small set of shared computational components rather than one separate strategy per person. Computational cognitive models have typically fixed a psychological architecture first and treated individual differences as variation inside a predefined parameterization, or examined them only after training. On the first Bernoulli Clicks Task dataset, held-out prediction reached its lowest validation loss at two dimensions each for temporal weighting and history dependence, yielding a six-dimensional representation of individual differences.

The model writes each participant's temporal weighting deviation and history kernel as a linear combination of shared components, with participant scores setting how strongly each component contributes. The low-rank factorization lets shared components and participant scores be learned jointly from trial-level choices across the population, instead of fitting each person separately and reducing dimensionality afterward. Temporal kernel parameters drop from 3100 to 346; the best population model has 990 parameters in total, about 6 per person, against 32 per person and 4960 for separate per-subject fits.

The two-dimensional temporal space continuously covers primacy, recency, middle-weighted, bookending and near-flat kernels, while the two-dimensional history space separates outcome-dependent from outcome-independent choice tendencies. These seemingly distinct heuristics had often been treated as independent behavioral strategies. In the newly collected replication dataset of 102 participants, validation loss again minimized at two dimensions each for temporal weighting and history dependence, and the same qualitative kernel shapes and strategy groupings reappeared.

The conclusions rest on auditory perceptual decision-making in the Bernoulli Clicks Task: 20 clicks within 1 second per trial, with the target side emitting a click with probability 0.55; 155 participants in the first dataset and 102 in the replication dataset, after excluding participants with accuracy below 60%.

In the replication dataset the side bias was numerically leftward overall but not significantly different from zero (W = 2347.00, p = .351), unlike the significant rightward bias seen in the first dataset; whether the two-dimensional temporal and history spaces generalize to other sensory modalities, continuous-time environments and different speed-accuracy regimes remains to be tested.