infrastructurefactbullishLinear transformers can reduce computational and memory complexity from quadratic to linear dependence on context length compared to softmax transformersMachine Learning (Statistics)27 Jul 2026http://arxiv.org/abs/2607.00479v1