Weird pattern in the ICLR papers I was assigned to review [D]

I was assigned 3 ICLR papers to review. I generally consider myself a pretty nice reviewer but...all 3 of them are slop. Unfiltered Claude slop. The language is so awful it is almost impossible to read. Unfortunately, I imagine this is common among "top" conference submissions now.

The reason I am posting here is because I noticed a strange similarity between all three papers. They all use very similar methods. They all start with the premise of deciding which of two predictors is better, and say that estimating predictor errors or scores individually can be difficult. So instead, they all propose to look at the difference between two errors or scores. *In all three cases*, when you look at the difference, some troublesome shared term cancels out, allowing you to derive a tighter bound on the difference. The authors then build a whole elaborate theoretical framework around this observation.

I know that paired comparisons and cancellation are standard techniques, but I don't think they are *so* common that I should get 3 separate submissions built around that same idea. Also, not only do all 3 submissions share the same core idea, but they all introduce it in the same way, use the same Claude-sounding prose, and include tons of additional proofs and experiments in a pretty similar way.

The coincidence is weird enough that I'm wondering if there is an autoresearch setup somewhere that has been spitting out papers of this sort. Has anyone else seen this idea as well?