RTPrune: Reading-Twice Inspired Token Pruning for Efficient DeepSeek-OCR Inference

2026-05-01

AI Summary

This work addresses the redundancy of visual tokens in DeepSeek-OCR, where existing pruning methods struggle to balance compression ratio and text fidelity. The authors propose RTPrune, a two-stage pruning approach that first retains high-norm critical tokens and then pairs and merges the remaining tokens based on optimal transport theory. A dynamic pruning ratio is introduced to adapt to varying text densities across documents. RTPrune further leverages the newly uncovered two-stage reading trajectory inherent in DeepSeek-OCR to design an efficient and accurate dynamic pruning mechanism. Evaluated on OmniDocBench, the method achieves 99.47% accuracy while retaining only 84.25% of visual tokens, yielding a 1.23× speedup in prefill latency and significantly outperforming existing vision-language model pruning techniques.

Abstract

DeepSeek-OCR leverages visual-text compression to reduce long-text processing costs and accelerate inference, yet visual tokens remain prone to redundant textual and structural information. Moreover, current token pruning methods for conventional vision-language models (VLMs) fail to preserve textual fidelity due to improper compression mechanisms. By analyzing the decoding process of DeepSeek-OCR, we find that a distinct two-stage reading trajectory: the model initially prioritizes the majority of high-norm tokens, then subsequently redistributes its attention to the remaining ones. Motivated by this insight, we propose RTPrune, a two-stage token pruning method tailored for DeepSeek-OCR. In the first stage, we prioritize high-norm visual tokens that capture salient textual and structural information. In the second stage, the remaining tokens are paired and merged based on optimal transport theory to achieve efficient feature aggregation. We further introduce a dynamic pruning ratio that adapts to token similarity and textual density for OCR tasks, enabling a better efficiency-accuracy trade-off. Extensive experiments demonstrate state-of-the-art performance, as evidenced by 99.47% accuracy and 1.23$\times$ faster prefill on OmniDocBench, achieved with 84.25% token retention when applied to DeepSeek-OCR-Large.

Problem

• token pruning
• textual fidelity
• visual-text compression
• redundant information
• OCR inference

Innovation

• token pruning
• two-stage reading
• optimal transport
• dynamic pruning ratio
• DeepSeek-OCR