I think an interesting avenue to explore is creating abstractions and analogies. If a model can take a novel situation and create an analogy to one that it is

I think an interesting avenue to explore is creating abstractions and analogies. If a model can take a novel situation and create an analogy to one that it is familiar with, it would expand its “reasoning” capabilities beyond its training data.