← back to paper
arxiv: 2605.00392 · 2 revisions
RTPrune: Reading-Twice Inspired Token Pruning for Efficient DeepSeek-OCR Inference