We study the ability of large-scale circulation models to reproduce extreme temperature events. To this end, we use a statistical clustering technique, Latent Dirichlet Allocation (LDA) to characterize sea-level pressure data over the north-Atlantic region. From the ERA5 reanalysis dataset, the method extracts a basis of interpretable objects at synoptic scale, that we call "motifs". Pressure data can be projected onto this basis, yielding motif weights that contain local information about the large-scale atmospheric circulation. We first examine how the weights statistics can be used to characterize extreme events in reanalysis data. We then compare the weights obtained from reanalysis data with those obtained from runs from four CMIP6 models. This allows us to quantify errors on each localized circulation pattern and identify model-agnostic and model-specific errors. On average, large-scale circulation is well predicted by all models, but model errors are increased for extreme events such as heatwaves and cold spells. A significant source of error was found to be associated with Mediterranean motifs for all models in all cases. Each model run can be characterized by a dynamic error associated with the global circulation pattern and a thermodynamic error associated with the predicted temperature. In the general case, this two-dimensional characterization is sufficient to discriminate between models. This remains possible in the cold spell case despite higher internal model variability, while all models perform similarly on heatwaves. The detailed characterization provided by LDA analysis is therefore well suited for model preselection for the study of extreme events.