infrastructurefactneutralDynamic sparsity methods improve quality retention by adapting computation to individual inputs, but their practical acceleration under real-world inference scenarios remains challengingMachine Learning02 Aug 2026http://arxiv.org/abs/2607.28418v1