一作NeurIPS 2026PosterStructural Self-Teaching for Compositional Generation in Unified Multimodal ModelsJaeyoo Park · Bohyung Han多模态长上下文与架构OpenReview