Baidu releases Unlimited-OCR 2 with Sparse Reference Sliding Window Attention for unlimited-horizon parsing

·2026.10.27 17:54

Key point

The model compresses attention to a fixed 512-token reference set to eliminate quadratic scaling and memory bottlenecks in long documents.

Baidu releases Unlimited-OCR 2, introducing Sparse Reference Sliding Window Attention (SR-SWA) to enable unlimited-horizon parsing within a finite context window.
SR-SWA compresses text-visual attention interactions from ultra-long sequences into a fixed-size set of working references, specifically 512 tokens.
This mechanism bounds computational complexity, keeping per-step decoding latency and memory footprint nearly constant regardless of document length.
The model achieves state-of-the-art performance on dozens-of-pages benchmarks like MPDocBench and extreme hundred-page benchmarks like MarathonDocBench.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.