{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:MHQFABVDPDJ25UCI6FPPO2QXTD","short_pith_number":"pith:MHQFABVD","schema_version":"1.0","canonical_sha256":"61e05006a378d3aed048f15ef76a1798fae4e7f6a4698c1b62fcea3962ec2680","source":{"kind":"arxiv","id":"2410.04417","version":4},"attestation_state":"computed","paper":{"title":"SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"SparseVLM prunes visual tokens in VLMs using text attention scores without any training or added parameters.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Chun-Kai Fan, Denis Gudovskiy, Junpeng Ma, Kuan Cheng, Kurt Keutzer, Shanghang Zhang, Tao Huang, Tomoyuki Okuno, Wenzhao Zheng, Yohei Nakata, Yuan Zhang","submitted_at":"2024-10-06T09:18:04Z","abstract_excerpt":"In vision-language models (VLMs), visual tokens usually bear a significant amount of computational overhead despite sparsity of information in them when compared to text tokens. To address this, most existing methods learn a network to prune redundant visual tokens using certain training data. Differently, we propose a text-guided training-free token optimization mechanism dubbed SparseVLM that eliminates the need of extra parameters or fine-tuning costs. Given that visual tokens complement text tokens in VLM's linguistic reasoning, we select relevant text tokens to rate the significance of vi"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":true,"formal_links_present":true},"canonical_record":{"source":{"id":"2410.04417","kind":"arxiv","version":4},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.CV","submitted_at":"2024-10-06T09:18:04Z","cross_cats_sorted":[],"title_canon_sha256":"5a5b188732cf551e1c8e59a062df3da8b29f8ba0ba1d9a822d2800cd56f8afd6","abstract_canon_sha256":"7f4d6f8fdd0d8c4f7bdd63f53aeab9330445dea86771847def61af753e0e9484"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:38:52.301150Z","signature_b64":"BJKAVmjjmJXhZh+KVg5lC4AFZJ1JX19hVGUcVg3BuWtXjD7P6DwPGIATXGKVlvCAlk+y9zQqxUMXXU8ZVu1+Cw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"61e05006a378d3aed048f15ef76a1798fae4e7f6a4698c1b62fcea3962ec2680","last_reissued_at":"2026-05-17T23:38:52.300692Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:38:52.300692Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"SparseVLM prunes visual tokens in VLMs using text attention scores without any training or added parameters.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Chun-Kai Fan, Denis Gudovskiy, Junpeng Ma, Kuan Cheng, Kurt Keutzer, Shanghang Zhang, Tao Huang, Tomoyuki Okuno, Wenzhao Zheng, Yohei Nakata, Yuan Zhang","submitted_at":"2024-10-06T09:18:04Z","abstract_excerpt":"In vision-language models (VLMs), visual tokens usually bear a significant amount of computational overhead despite sparsity of information in them when compared to text tokens. To address this, most existing methods learn a network to prune redundant visual tokens using certain training data. Differently, we propose a text-guided training-free token optimization mechanism dubbed SparseVLM that eliminates the need of extra parameters or fine-tuning costs. Given that visual tokens complement text tokens in VLM's linguistic reasoning, we select relevant text tokens to rate the significance of vi"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"SparseVLM increases the efficiency of various VLMs in a number of image and video understanding tasks. For example, LLaVA when equipped with SparseVLM achieves 54% reduction in FLOPs, 37% decrease in CUDA latency while maintaining 97% of its original accuracy.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That self-attention scores between selected text tokens and visual tokens reliably identify which visual tokens can be pruned or recycled without losing task-critical information.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"SparseVLM uses text-guided attention to prune and recycle visual tokens in VLMs, delivering 54% FLOPs reduction and 37% lower latency with 97% accuracy retention on LLaVA.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"SparseVLM prunes visual tokens in VLMs using text attention scores without any training or added parameters.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"f5f7695a3f895ec4f477af7edd6fda775677f5cf825fee44489a8d8b021ab3d8"},"source":{"id":"2410.04417","kind":"arxiv","version":4},"verdict":{"id":"6bba3ee3-d597-404a-8061-2f18b500a1d1","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T14:51:52.984961Z","strongest_claim":"SparseVLM increases the efficiency of various VLMs in a number of image and video understanding tasks. For example, LLaVA when equipped with SparseVLM achieves 54% reduction in FLOPs, 37% decrease in CUDA latency while maintaining 97% of its original accuracy.","one_line_summary":"SparseVLM uses text-guided attention to prune and recycle visual tokens in VLMs, delivering 54% FLOPs reduction and 37% lower latency with 97% accuracy retention on LLaVA.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That self-attention scores between selected text tokens and visual tokens reliably identify which visual tokens can be pruned or recycled without losing task-critical information.","pith_extraction_headline":"SparseVLM prunes visual tokens in VLMs using text attention scores without any training or added parameters."},"references":{"count":113,"sample":[{"doi":"","year":2022,"title":"Flamingo: a visual language model for few-shot learning","work_id":"fa381b5e-5991-4bbf-a49a-2e90138d814f","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond","work_id":"cbc2bb21-b6bb-46c0-80bf-107e195ffe10","ref_index":2,"cited_arxiv_id":"2308.12966","is_internal_anchor":true},{"doi":"","year":2023,"title":"Token merging: Your vit but faster","work_id":"671514fa-294d-462f-b530-bd25a0b38101","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2020,"title":"D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al","work_id":"9c4d1f20-8715-4aaa-9bb1-9420a625a69d","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Cai, M., Yang, J., Gao, J., and Lee, Y. J. Matryoshka multimodal models. In International Conference on Learning Representations, 2025","work_id":"9dedb931-d4a0-4bbe-8674-ef4b4d38a5ed","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":113,"snapshot_sha256":"e1a38f11346fc40939c1883d42f7127bb0b240b3dbe26bec2e7c50b547b9be06","internal_anchors":12},"formal_canon":{"evidence_count":2,"snapshot_sha256":"3e88488bc8d9ffba410b08778160f2c70a843dff80be28da54870cd7ca6e3d1d"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2410.04417","created_at":"2026-05-17T23:38:52.300759+00:00"},{"alias_kind":"arxiv_version","alias_value":"2410.04417v4","created_at":"2026-05-17T23:38:52.300759+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2410.04417","created_at":"2026-05-17T23:38:52.300759+00:00"},{"alias_kind":"pith_short_12","alias_value":"MHQFABVDPDJ2","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"MHQFABVDPDJ25UCI","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"MHQFABVD","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":31,"internal_anchor_count":31,"sample":[{"citing_arxiv_id":"2503.14075","citing_title":"Growing a Multi-head Twig via Distillation and Reinforcement Learning to Accelerate Large Vision-Language Models","ref_index":65,"is_internal_anchor":true},{"citing_arxiv_id":"2605.22372","citing_title":"ASAP: Attention Sink Anchored Pruning","ref_index":9,"is_internal_anchor":true},{"citing_arxiv_id":"2605.20950","citing_title":"Focus-then-Context: Subject-Centric Progressive Visual Token Reduction for Vision-Language Models","ref_index":5,"is_internal_anchor":true},{"citing_arxiv_id":"2605.15621","citing_title":"LRCP: Low-Rank Compressibility Guided Visual Token Pruning for Efficient LVLMs","ref_index":25,"is_internal_anchor":true},{"citing_arxiv_id":"2605.19218","citing_title":"Rotation-Aligned Key Channel Pruning for Efficient Vision-Language Model Inference","ref_index":35,"is_internal_anchor":true},{"citing_arxiv_id":"2605.19322","citing_title":"DynaTok: Temporally Adaptive and Positional Bias-Aware Token Compression for Video-LLMs","ref_index":34,"is_internal_anchor":true},{"citing_arxiv_id":"2511.18960","citing_title":"AVA-VLA: Improving Vision-Language-Action models with Active Visual Attention","ref_index":54,"is_internal_anchor":true},{"citing_arxiv_id":"2511.22663","citing_title":"AIA: Rethinking Architecture Decoupling Strategy In Unified Multimodal Model","ref_index":51,"is_internal_anchor":true},{"citing_arxiv_id":"2512.19219","citing_title":"Selective LoRA for Visual Tokens and Attention Heads","ref_index":30,"is_internal_anchor":true},{"citing_arxiv_id":"2601.01322","citing_title":"LinMU: Multimodal Understanding Made Linear","ref_index":22,"is_internal_anchor":true},{"citing_arxiv_id":"2603.01400","citing_title":"Token Reduction via Local and Global Contexts Optimization for Efficient Video Large Language Models","ref_index":78,"is_internal_anchor":true},{"citing_arxiv_id":"2410.17247","citing_title":"PyramidDrop: Accelerating Your Large Vision-Language Models via Pyramid Visual Redundancy Reduction","ref_index":54,"is_internal_anchor":true},{"citing_arxiv_id":"2603.22911","citing_title":"ForestPrune: High-ratio Visual Token Compression for Video Multimodal Large Language Models via Spatial-Temporal Forest Modeling","ref_index":64,"is_internal_anchor":true},{"citing_arxiv_id":"2603.27960","citing_title":"Towards Efficient Large Vision-Language Models: A Comprehensive Survey on Inference Strategies","ref_index":77,"is_internal_anchor":true},{"citing_arxiv_id":"2605.13548","citing_title":"AttenA+: Rectifying Action Inequality in Robotic Foundation Models","ref_index":27,"is_internal_anchor":true},{"citing_arxiv_id":"2604.04055","citing_title":"DINO-VO: Learning Where to Focus for Enhanced State Estimation","ref_index":68,"is_internal_anchor":true},{"citing_arxiv_id":"2605.09649","citing_title":"Make Each Token Count: Towards Improving Long-Context Performance with KV Cache Eviction","ref_index":34,"is_internal_anchor":true},{"citing_arxiv_id":"2605.08985","citing_title":"LLaVA-UHD v4: What Makes Efficient Visual Encoding in MLLMs?","ref_index":53,"is_internal_anchor":true},{"citing_arxiv_id":"2605.03351","citing_title":"VLMaxxing through FrameMogging Training-Free Anti-Recomputation for Video Vision-Language Models","ref_index":31,"is_internal_anchor":true},{"citing_arxiv_id":"2604.24447","citing_title":"Characterizing Vision-Language-Action Models across XPUs: Constraints and Acceleration for On-Robot Deployment","ref_index":29,"is_internal_anchor":true},{"citing_arxiv_id":"2605.05848","citing_title":"VideoRouter: Query-Adaptive Dual Routing for Efficient Long-Video Understanding","ref_index":10,"is_internal_anchor":true},{"citing_arxiv_id":"2605.04227","citing_title":"Pro$^2$Assist: Continuous Step-Aware Proactive Assistance with Multimodal Egocentric Perception for Long-Horizon Procedural Tasks","ref_index":86,"is_internal_anchor":true},{"citing_arxiv_id":"2605.00392","citing_title":"RTPrune: Reading-Twice Inspired Token Pruning for Efficient DeepSeek-OCR Inference","ref_index":3,"is_internal_anchor":true},{"citing_arxiv_id":"2604.12358","citing_title":"Why and When Visual Token Pruning Fails? A Study on Relevant Visual Information Shift in MLLMs Decoding","ref_index":53,"is_internal_anchor":true},{"citing_arxiv_id":"2604.11240","citing_title":"Decoupled Similarity for Task-Aware Token Pruning in Large Vision-Language Models","ref_index":48,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/MHQFABVDPDJ25UCI6FPPO2QXTD","json":"https://pith.science/pith/MHQFABVDPDJ25UCI6FPPO2QXTD.json","graph_json":"https://pith.science/api/pith-number/MHQFABVDPDJ25UCI6FPPO2QXTD/graph.json","events_json":"https://pith.science/api/pith-number/MHQFABVDPDJ25UCI6FPPO2QXTD/events.json","paper":"https://pith.science/paper/MHQFABVD"},"agent_actions":{"view_html":"https://pith.science/pith/MHQFABVDPDJ25UCI6FPPO2QXTD","download_json":"https://pith.science/pith/MHQFABVDPDJ25UCI6FPPO2QXTD.json","view_paper":"https://pith.science/paper/MHQFABVD","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2410.04417&json=true","fetch_graph":"https://pith.science/api/pith-number/MHQFABVDPDJ25UCI6FPPO2QXTD/graph.json","fetch_events":"https://pith.science/api/pith-number/MHQFABVDPDJ25UCI6FPPO2QXTD/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/MHQFABVDPDJ25UCI6FPPO2QXTD/action/timestamp_anchor","attest_storage":"https://pith.science/pith/MHQFABVDPDJ25UCI6FPPO2QXTD/action/storage_attestation","attest_author":"https://pith.science/pith/MHQFABVDPDJ25UCI6FPPO2QXTD/action/author_attestation","sign_citation":"https://pith.science/pith/MHQFABVDPDJ25UCI6FPPO2QXTD/action/citation_signature","submit_replication":"https://pith.science/pith/MHQFABVDPDJ25UCI6FPPO2QXTD/action/replication_record"}},"created_at":"2026-05-17T23:38:52.300759+00:00","updated_at":"2026-05-17T23:38:52.300759+00:00"}