>It is, and it should, be able to find correlations across data no matter how the input is arranged. That doesn't mean it doesn't favor certain arrangements.

>It is, and it should, be able to find correlations across data no matter how the input is arranged. That doesn't mean it doesn't favor certain arrangements. https://arxiv.org/abs/2305.18354 https://arxiv.org/abs/2305.18354 Linear is in fact better for this benchmark