{"paper":{"title":"RTPrune: Reading-Twice Inspired Token Pruning for Efficient DeepSeek-OCR Inference","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"RTPrune applies a two-stage pruning process to DeepSeek-OCR that first keeps high-norm visual tokens and then merges the rest with optimal transport to cut inference time while preserving OCR accuracy.","cross_cats":["cs.LG"],"primary_cat":"cs.CV","authors_text":"Ben Wan, Jia Wang, Tongxuan Liu, Weizhe Huang, Yan Feng, Yuting Zeng, Zihan Tang","submitted_at":"2026-05-01T04:30:16Z","abstract_excerpt":"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"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Extensive experiments demonstrate state-of-the-art performance, as evidenced by 99.47% accuracy and 1.23× faster prefill on OmniDocBench, achieved with 84.25% token retention when applied to DeepSeek-OCR-Large.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The observed two-stage reading trajectory in DeepSeek-OCR is stable and general enough that pruning high-norm tokens first followed by optimal-transport merging of the rest will preserve textual fidelity across OCR tasks.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"RTPrune delivers 99.47% accuracy and 1.23x faster prefill on OmniDocBench for DeepSeek-OCR-Large by retaining only 84.25% of tokens through a reading-twice inspired two-stage pruning process.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"RTPrune applies a two-stage pruning process to DeepSeek-OCR that first keeps high-norm visual tokens and then merges the rest with optimal transport to cut inference time while preserving OCR accuracy.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"0d9d8ba1dc9964d0b82726a30d44176d8ff830b52192238e23995f1187f20bbf"},"source":{"id":"2605.00392","kind":"arxiv","version":3},"verdict":{"id":"9ad105ae-5aa3-48ec-8548-c6580e248ac8","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-12T03:07:12.746382Z","strongest_claim":"Extensive experiments demonstrate state-of-the-art performance, as evidenced by 99.47% accuracy and 1.23× faster prefill on OmniDocBench, achieved with 84.25% token retention when applied to DeepSeek-OCR-Large.","one_line_summary":"RTPrune delivers 99.47% accuracy and 1.23x faster prefill on OmniDocBench for DeepSeek-OCR-Large by retaining only 84.25% of tokens through a reading-twice inspired two-stage pruning process.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The observed two-stage reading trajectory in DeepSeek-OCR is stable and general enough that pruning high-norm tokens first followed by optimal-transport merging of the rest will preserve textual fidelity across OCR tasks.","pith_extraction_headline":"RTPrune applies a two-stage pruning process to DeepSeek-OCR that first keeps high-norm visual tokens and then merges the rest with optimal transport to cut inference time while preserving OCR accuracy."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.00392/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-20T19:43:24.710508Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T18:12:25.211188Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"dbc464ed383eb8a369f4a6f19ae2bc067297355c3aa12322dc69e210380a06c9"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}