Abstract page for arXiv paper 2611.17754: See Every Block Once: Variance-Reduced Post-Training for Masked Diffusion Language Models