Dehaene & Changeux (1991) made a neural-network model:
The coding units are clusters of neurons organized in layers, or assemblies. A sensonmotor loop enables the network to sort the input cards according to several criteria (color, form, etc.). A higher-level assembly of rule-coding clusters codes for the currently tested rule, which shifts when negative reward is received. Internal testing of the possible rules, analogous to a reasoning process, also occurs, by means of an endogenous autoevaluation loop. When lesioned, the model reproduces the behavior of frontal lobe patients.
Parks et al. (1992) extended the previous neural models of WCST to take into account verbal-fluency.
Amos (2000) built a neural-network model that helps distinguish between the sort of errors made by patients with schizophrenia, Parkinson's disease, and Huntington's disease. He relates the model to neuroanatomy and says:
The model also made specific, empirically falsifiable predictions that can be used to explore the utility of these putative mechanisms of information processing in the frontal cortex and basal ganglia.
Monchi et al. (2000) followed a similar approach as Amos. Their model suggested different impairments in Parkinson’s disease and schizophrenic patients, and made specific predictions of what would be observed in fMRI scans. They tested this prediction in their paper.
By 2005, the WCST had become a benchmark for more general models. Rougier et al. (2005) made a general model of the prefrontal cortex based on general neurobiological principles as opposed to symbolic approaches. They model learnt from experience and generalized to novel tasks. They tested it on WCST and the Stroop task data for typical and frontally damaged subjects.
Kaplan et al. (2006) use a neural-network approach with two parts: a Hopfield net serves as the working memory and a Hamming block as a hypothesis generator.
Bishara et al. (2010) use a more ACT-R-like (although not actual ACT-R) model with probabilistic loading on symbolic rules. The resulting sequential learning model is used to identify specific process that subjects might be struggling with. It is suppose to help diagnosis in a clinical setting, and they test it on substance dependent individuals.
Rigotti et al. (2010) created a network of randomly connected neurons that could solve the task. The key point was that the random connections induced mixed selectivity, which would solve the task with high probability.