infrastructurecritiquebearishExisting LLM acceleration methods rely on task-specific fine-tuning or training from scratch, which increases adaptation cost and limits cross-task usabilityMachine Learning28 Jul 2026http://arxiv.org/abs/2607.21291v1