Yeah you absolutely can for GPCMs, selecting the most "informative items" means that you prioritizing those that provides most precision in estimating θ (for the relevant range of interest).

However, this also means you're potentially discarding items that actually capture important aspects of a construct that theoretically is needed in order to fully interpret the construct, i.e., content validity is potentially reduced. But more importantly, the item information function depends on the θ distribution, so items you find to be non-relevant in your population could turn out to be pretty important in other populations.

So yes, it is a valid approach, but these are some "drawbacks" you absolutely should address as weaknesses or limitations

All items with the same number of categories have the same total amount of information. A graph of the information function shows how the item's information is spread along the latent variable. You can choose items whose information peaks at important places, such as cut-points, but be careful to keep content and construct validity of the set of items, even if that means retaining items with less information where you want information to be at a maximum. There are two approaches to test information (= sum of item information): 1. it should peak at cut-points (pass-fail points), so that person estimates are more precise at cut-points. 2. it should be about level across the ability range, so that everyone is estimated about equally precisely.