{"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"}