SDAR is a synergistic diffusion-autoregressive paradigm that combines the training efficiency of autoregressive models with the parallel inference capabilities of diffusion models. Instead of expensive end-to-end diffusion training, SDAR transforms a well-trained autoregressive (AR) model into a block-wise diffusion model through simple, data-efficient adaptation. During inference, SDAR autoregressively generates sequences across blocks to ensure global consistency, while simultaneously decoding all tokens within each block in parallel through a discrete diffusion process. AR models are significantly more computationally efficient than masked diffusion models, and based on this, SDAR achieves an efficient AR-to-diffusion transformation at minimal cost, enabling parallel generation while maintaining AR-level performance. Large-scale model studies demonstrate that SDAR is robust to block size and decoding thresholds, delivering significant speedups without loss of accuracy. SDAR also demonstrates enhanced inference capability and domain adaptability. The 30B MoE model outperforms AR models on demanding scientific inference benchmarks such as GPQA and ChemBench, and further improves with test time scaling methods such as majority voting and pass@k.
