More than Just Measurement Models? Processing Insights from Population Coding Models of Visual Working Memory Retrieval

Continuous memory retrieval tasks provide a precise measure of visual working memory (VWM), but the mechanisms underlying their heavy-tailed error distributions remain debated. Recent population coding models (PCMs), including the neural resource model , target confusability competition model , and signal discrimination model , explain such errors as the outcome of noisy signals distributed over feature space. These models share a common tuning–noise–decoding structure but differ in their tuning functions, noise distributions, decoding rules, and process-level interpretations. Across two colour reproduction experiments, we tested how memory load and stimulus identifiability are expressed within this shared framework. In Experiment 1, participants remembered colour arrays varying in set size and stimulus saturation. In Experiment 2, we factorially manipulated stimulus and response-wheel saturation while varying set size. We first compared models formed by crossing von Mises and Laplace-like tuning functions with Normal, Gumbel, and Poisson noise distributions. We then tested whether set-size effects and saturation effects were selectively identified in the amplitude parameter \:γ\: and dispersion parameter \:κ\: , respectively. Finally, we evaluated whether the set-size effect followed a sample-size model constraint, or its more flexible power-law generalisation. We found that the preferred tuning function varied across individuals, whereas Normal and Gumbel noise generally provided better fits than Poisson noise. Model comparisons supported separable effects of memory load and stimulus identifiability, captured selectively by changes in \:γ\: and \:κ\: . Across both experiments, \:γ\: decreased closely following a square-root law with set size, supporting the prediction of sample-size model. Our findings provide direct empirical evidence supporting that PCMs do more than measure continuous error distributions: their parameters provide identifiable links between empirical VWM performances and fundamental cognitive constructs, including memory capacity limits and the structure of mental stimulus representations.