How to Train Large Language Models to Be More Effective

Large language models are getting good at things like block world and maybe even more complicated reasoning tasks where you have to figure out who's someone's cousin based on some logic. They're not necessarily a superhuman level yet could it just be we need more time and practice and the reasoning will be trained in? I cannot say no because those those big models do not stop to amaze me they are still lacking especially in the like visual the puzzles department so I saw papers I don't remember them from the top of my head but yeah maybe we will get there.