Deep convolutional neural network (DNN)- based models have emerged as our leading hypotheses of human vision. Here we describe, and expand upon, our latest effort to use DNN models of brain regions to explain key results from previous cognitive neuroscience and psychology experiments. Many stimuli in these prior experiments were highly manipulated (e.g. scrambled body parts, face parts, re-arranged spatial positions) often outside the domain of natural stimuli. These results can therefore be considered as tests of model generalization beyond naturalistic stimuli. We first performed these tests on the fusiform face area (FFA), parahippocampal place area (PPA) and the extrastriate body area (EBA). Our previous results (presented in VSS2022) showed that our fROI-level models recapitulate several key results from prior studies. We also observed that models did not perform as well on non-naturalistic stimuli. Here we extend our model evaluation metrics in two ways. First, we replicated findings from the original paper on the EBA (Downing et al 2001) that the EBA responds as highly to line drawings of bodies, and symbolic stick figures as to natural images of bodies (and not to control conditions like faces and objects). Second, we find that none of the computational models explain this pattern of observed responses, though models trained with language-based supervision (like CLIP) do better than other models. Together, our results on symbolic body images expose the bounds of current computational models. This progress was made possible only because of fROI-level modeling procedures, and opens up new ways to understand the power and limitations of current models and test novel hypotheses completely in-silico.