A real test would be to invent a framework and see if it can do anything on the new framework. Otherwise it’s just a search engine that amalgamates pieces
Is it context size, training, number of iterations, number of agents? That's what's being tried to improve the results.