Prefix-Adaptive Block Diffusion for Efficient Document Recognition

In three linesPA-BDM improves Block Diffusion Models for document recognition by replacing bidirectional denoising with causal prefix-to-suffix denoising. Using Confidence-gated Structural Loss and Progressive Prefix Commitment, the 3B model achieves 71.6% higher inference throughput than MinerU-Diffusion 2.5B.

Summary generated by Claude — human-verified