The reason I draw this conclusion is because if you'll look at planning engines you'll see that they are operating on the symbolic knowledge and reasoning is also symbolic, given that LLMs use heuristics, it is unlikely pure transformer-based model will automatically develop that sort of capabilities.

The fact that LLMs are not learning the underlying logic, instead using heuristics, makes it highly unlikely that it is not abstracting properly to do reasoning. To do skill acquisition, the system should rely more towards abstraction rather than heuristics.