LaViDa-O, our unified diffusion model for multimodal understanding and generation, appears at ICLR 2026.