RTPrune: Reading-Twice Inspired Token Pruning for Efficient DeepSeek-OCR Inference
> Paper: RTPrune: Reading-Twice Inspired Token Pruning for Efficient DeepSeek-OCR Inference > Authors: Ben Wan, Yan Feng, Zihan Tang, Weizhe Huang > arXiv: 2605.00392 | 2026-04-29
The OCR Dilemma: Too Many Visual Tokens
When running DeepSeek-OCR on a scanned book:
- Each page image produces a large number of visual tokens
- Many tokens are redundant:
- Blank regions
- Repeated text structures
- Decorative elements
- Processing is slow
- Memory usage is high
- Designed for general vision-language models, not OCR
- Pruning loses text details
- OCR accuracy drops after pruning
- First pass: quick scan to grasp structure
- Second pass: careful reading of important parts
- Not every word gets equal attention
- First pass: quickly scan all tokens and identify important regions
- Second pass: process only the important tokens, ignore redundancy
- Understands text structure
- Preserves text regions and key layout elements
- Prunes whitespace, decorations, and repeated structures
- Traditional pruning loses text details
- RTPrune specifically protects textual information
- OCR accuracy does not degrade after pruning
- Fewer tokens processed
- Faster inference
- Lower memory footprint
- Traditional methods: read the entire book word by word
- RTPrune: glance at the table of contents and chapter headings, then read the important paragraphs
- Faster, without missing key information
- All tokens treated equally, wasting compute on redundancy
- Documents have structured layouts (headings, paragraphs, tables) that generic pruning does not understand
- Structure awareness: the first pass understands document structure; the second focuses on what matters
- Text protection: it knows what is "text" versus "decoration," enabling precise retention and pruning
- Efficiency–quality balance: major speedup with no accuracy loss
Problems with existing pruning methods:
What's needed: token pruning designed specifically for OCR.
RTPrune: The Reading-Twice Inspiration
The paper draws on how humans read long documents:
Technical approach:
1. Two-pass reading mechanism
2. OCR-aware pruning
3. Text fidelity
4. Efficiency gains
Analogy:
Why "Reading Twice" Beats "Reading Once"
Problems with single-pass processing:
Advantages of the two-pass approach:
Takeaways
If you are optimizing document processing or OCR systems, ask:
1. "Is my system processing too many redundant tokens?" 2. "Does the OCR scenario call for a specific pruning strategy?" 3. "Could a two-pass mechanism improve processing efficiency?" 4. "Does pruning protect critical text information?"
RTPrune reminds us: in OCR, the fastest processing is not processing everything, but processing only what matters.
When DeepSeek-OCR learns to "read twice"—once for structure, once for the key content—it becomes not only faster but smarter. In the future of document intelligence, selective attention is the key to efficiency. The best readers are not the fastest; they are the ones who focus selectively.