Unlimited OCR Works 2
The authors report up to 7× faster decoding steps on hundred-page documents while keeping per-step computation constant as documents grow.
Unlimited OCR 2 caps the reference tokens active in each decoder attention module using Sparse Reference Sliding Window Attention. This addresses the rising decoding cost caused by expanding visual context in its predecessor. At the hundred-page scale, the reported per-step speedups are 7.0× over Full Attention and 5.6× over R-SWA. The authors also report state-of-the-art end-to-end results on MPDocBench and their new hundred-page parsing benchmark, MarathonDocBench.
ArXiv · AI/CL/LG's note
Unlimited OCR 2 caps the reference tokens active in each decoder attention module using Sparse Reference Sliding Window Attention. This addresses the rising decoding cost caused by expanding visual context in its predecessor. At the hundred-page scale, the reported per-step speedups are 7.0× over Full Attention and 5.6× over R-SWA. The authors also report state-of-the-art end-to-end results on MPDocBench and their new hundred-page parsing benchmark, MarathonDocBench.
ArXiv · AI/CL/LG's note
score 5