Sure. Not in the sense of "here’s a proof that a perceptron cannot learn XOR", but at least something that would hint that there are qualitative gaps in the exp

Sure. Not in the sense of "here’s a proof that a perceptron cannot learn XOR", but at least something that would hint that there are qualitative gaps in the expressive power of DLNNs, or even LLMs specifically.