1) I think the figure for showing convergence of parameter was redundant, in the other words, those figures needn't to be shown because the convergence of parameters can be simply reported in the paragraph. And in this article, those figures didn't explain any meaning about the DGM model, that is author also think the convergence figure was not important.
2) I am always curious that can it be applied in the practice when we adjusted cheaters' score just according to their response?
3) The author used BILOG-MG to calibrate latent trait based on the unexposed items as the true ability. Then he computed the correlation between BILOG-MG and DGM based on all items. However, as what he mentioned in model, the DGM just using unexposed items should be equal to Rasch model. So I think the comparison of BILOG-MG and DGM was actually the comparison of two identical model. The difference was only caused by the size of data, one using unexposed items and the other using all items.
4) I more encourage to use mixture model to express the response time (RT) in computerized test, and the cheating effect can correlate with the RT parameters.