Taking a look at a paper I have that went up this month and another that went up before the dec cutoff on ar5iv, they look 90% OK! Figures with side-by-side plots and algorithm environments are the common culprit for being broken though. Particularly in figures, it seems like the width argument isn't being interpreted correctly.
Interestingly this review paper seems to have their side by side figures intact (e.g. fig 2 fig 4). Maybe it's because he used a subfigure like environment (judging by the subcaptions)?