Diffusion Drafts, AR Verifies: Accelerating Document OCR with Self-Speculative Decoding
GravityOCR speeds document OCR by drafting several tokens in parallel, then verifying them with an autoregressive path before committing.
The paper says this avoids using a separate drafting network while reducing errors from parallel token prediction. Its shared AR path also supports GRPO training with OCR rewards for sequence and structure quality. On OmniDocBench v1.6, AR-path GRPO raises the Overall score from 94.92 to 95.16, while the final model stays near GLM-OCR’s 95.48. In SGLang serving, it averages 9.7 committed tokens per forward pass, with a 3.94x decode-only speedup on region crops and a 1.32x end-to-end page speedup over AR decoding. ArXiv · AI/CL/LG's note
The paper says this avoids using a separate drafting network while reducing errors from parallel token prediction. Its shared AR path also supports GRPO training with OCR rewards for sequence and structure quality. On OmniDocBench v1.6, AR-path GRPO raises the Overall score from 94.92 to 95.16, while the final model stays near GLM-OCR’s 95.48. In SGLang serving, it averages 9.7 committed tokens per forward pass, with a 3.94x decode-only speedup on region crops and a 1.32x end-to-end page speedup over AR decoding. ArXiv · AI/CL/LG's note
score 5