Diffusion Drafts, AR Verifies: Accelerating Document OCR with Self-Speculative Decoding

TL;DR - GravityOCR combines diffusion-based parallel drafting with autoregressive verification in one parameter-shared OCR model. It accelerates document decoding while retaining accuracy close to the original GLM-OCR model.

• The model drafts multiple tokens in parallel, then causally verifies them before commitment to reduce errors.
• Its autoregressive path supports GRPO using sequence- and structure-level OCR rewards without diffusion-trajectory likelihood estimation.
• On OmniDocBench v1.6, GRPO improved the Overall score from 94.92 to 95.16; the final model scored near GLM-OCR’s 95.48.
• In SGLang, GravityOCR committed 9.7 tokens per forward pass on average, delivering 3.94× decode-only and 1.32× end-to-end speedups over autoregressive decoding.