I think you are overthinking this. The ARC benchmark for fluid abstracting reasoning was made in 2019 and it still hasn't been 'solved'. So the goalposts aren't moving as much as you think they are.

LLMs or neural nets have never been good with out of distribution tasks.