multimodalfactbearishExisting SLM training approaches are difficult to scale because speech sequences are significantly longer than text sequencesComputation and Language27 Jul 2026http://arxiv.org/abs/2607.02214v1