I realize (having been back and forth on this) that the answer is a very strong yes.
LLMs are based on language. The concepts/"nodes" LLMs tie together are useful because humans have created these nodes over the entire history of language, culture, and science. Humans naturally over millennia have been filtering out useful concepts from useless ones, refining language to efficiently carry meaning.
If we broke down all abstractions to fundamentals, there would be billions of things that would need to be tied together to make any useful product, almost impossible. Abstractions are required for solving problems efficiently. A high level abstraction that fits well in many solutions is insanely powerful.
LLMs are not immune to this effect, they need abstractions so that they don't have to prove every single thing they do from first principles, they can instead lean on language to shortcut a ton of thinking.
So when a framework gives you useful high level structure, patterns, and abstractions, you can do a lot less reasoning, and work a lot more efficiently, to solve a lot more problems. You could have one word (e.g. controller) mean a whole ton of useful things that you otherwise would've had to describe with multiple words. You could have references to existing concepts (e.g. `String#camelize`) that you no longer need to expand and define each time.
Bottom line, abstractions let LLMs solve problems faster, more correctly, with less reasoning.
P.S. Language is all abstractions, and we can solve problems well because we created these effective abstractions. The question I have about LLMs going forward is whether they can create new useful abstractions (new words, concepts, categories) that are actually time-tested and proven to contribute to good problem solving, without human input/verification.