infrastructurefactbearishTransformer architecture incurs high inference costs for large language modelsMachine Learning28 Jul 2026http://arxiv.org/abs/2607.21291v1