Take a common word problem in a 5th grade math text book. Now, change as many words as possible; instead of two trains, make it two different animals; change the location to a rarely discussed town; etc. Even better, invent words/names to identify things.
Someone who has done a word problem like that will very likely recognize the logic, even if the setting is completely different.
A lot of LLMs do weird things on the question "A farmer needs to get a bag of grain across a river. He has a boat that can transport himself and the grain. How does he do this?"
(they often pattern-match on the farmer/grain/sheep/fox puzzle and start inventing pointless trips ("the farmer returns alone. Then, he crosses again.") in a way that a human wouldn't)
I have noted over my life that a lot of problems end up being a variation on solved problems from another more familiar domain but frustratingly take a long time to solve before realizing this was just like that thing you had already solved. Nevertheless, I do feel like humans do benefit from identifying meta patterns but as the chess example shows even we might be weak in unfamiliar areas.
Learn how to solve one problem and apply the approach, logic and patterns to different problems. In German that's called "Transferleistung" (roughly "transfer success") and a big thing at advanced schools. Or, at least my teacher friends never stop talking about it.
We get better at it over time, as probably most of us can attest.
Take a common word problem in a 5th grade math text book. Now, change as many words as possible; instead of two trains, make it two different animals; change the location to a rarely discussed town; etc. Even better, invent words/names to identify things.
Someone who has done a word problem like that will very likely recognize the logic, even if the setting is completely different.
Word tokenization alone should fail miserably.