multimodalfactbullishSynthetic domain-specific data combined with improved speech projection allows models to outperform previous best systems while being considerably more efficientComputation and Language27 Jul 2026http://arxiv.org/abs/2607.01960v1