Prefix-Adaptive Block Diffusion for Efficient Document Recognition

2026-05-16

AI Summary

Existing block diffusion models suffer from limited parallelism and cache latency due to fixed block boundaries, and their inconsistent information flow—bidirectional denoising within blocks versus autoregressive propagation across blocks—degrades performance on structure-sensitive document recognition tasks. To address these issues, this work proposes the Prefix-Adaptive Block Diffusion Model (PA-BDM), which reformulates intra-block denoising as a causal process from prefix to suffix and treats block size as a dynamic candidate range rather than a fixed unit. PA-BDM integrates a confidence-gated structural loss (CSL) and a progressive prefix commitment (PPC) mechanism to enable efficient parallel decoding while preserving recognition accuracy. With 3B parameters, PA-BDM outperforms current methods across multiple document recognition benchmarks and achieves a 71.6% higher inference throughput compared to the 2.5B-parameter MinerU-Diffusion model.

Abstract

Block Diffusion Models (BDMs) support parallel generation, flexible-length output, and KV caching, making them promising for efficient document parsing. However, existing BDMs bind denoising and cache commitment to fixed block boundaries: parallelism shrinks during intra-block denoising, while generated tokens cannot be cached until the whole block is completed. Moreover, intra-block bidirectional denoising conflicts with inter-block autoregression, creating inconsistent information flow that can challenge structure-sensitive recognition. We propose the Prefix-Adaptive Block Diffusion Model (PA-BDM), which replaces intra-block bidirectional denoising with causal denoising from prefix to suffix and treats the block size as a maximum candidate range rather than a fixed commitment unit. PA-BDM uses Confidence-gated Structural Loss (CSL) to build low-entropy prefixes before extending training to longer continuations. During inference, Progressive Prefix Commitment (PPC) then dynamically commits the longest reliable prefix into the KV cache and resets the next candidate range from the updated prefix, restoring a large parallel decoding space at each step. Experiments show that the 3B PA-BDM achieves higher recognition scores on several benchmarks and improves inference throughput by 71.6\% over the 2.5B MinerU-Diffusion.

Problem

• Block Diffusion Models
• document recognition
• parallel generation
• KV caching
• information flow inconsistency

Innovation

• Prefix-Adaptive Block Diffusion
• Causal Denoising
• Progressive Prefix Commitment
• Confidence-gated Structural Loss
• KV Caching