For climate models to continue improving, we need to uncover as many discrepancies they have with reality as possible. In particular, evaluating the representation of extreme events is important but challenging owing to their rarity. Here, we study how general circulation models reproduce large-scale atmospheric circulation associated with extreme temperature events. To this end, we apply Latent Dirichlet Allocation (LDA), a dimensionality reduction method, to a set of sea-level pressure ERA5 maps over the north-Atlantic region. LDA provides a basis of sparse latent modes called ``motifs'' that consist of localized objects at synoptic scale. Any pressure map can be approximated by a generally sparse combination of motifs, whose coefficients are called the weights, containing local information about large-scale circulation. Weights statistics can be used to locally characterize circulation patterns, in general and during extreme events, allowing for detailed comparison of datasets. For four CMIP6 models and reanalysis, we quantify local circulation errors and identify model-agnostic and model-specific biases. On average, large-scale circulation is well predicted by all models, but model errors are increased for heatwaves and cold spells. Significant errors were found to be associated with Mediterranean motifs for all models in all cases. In addition, the combination of motif and temperature error can discriminate between models in the general and cold spell cases, while models perform similarly on heatwaves. The sparse characterization provided by LDA analysis is therefore well suited for model preselection for the study of extreme events.