{ "schema_version": 1, "last_updated": "2026-10-20", "source": "README.md", "repository": "https://github.com/SuDIS-ZJU/Efficient-LVLMs-Inference", "stats": { "papers": 105, "with_code": 21, "with_project_page": 4, "with_resources": 25, "stages": { "Decoding": 32, "Encoding": 39, "Prefilling": 36 } }, "papers": [ { "title": "3DZip: Spatial-Aware Feature Diversity-Guided Token Compression for 3D Question Answering", "paper_url": "https://arxiv.org/abs/2608.01185", "venue": "ECCV 2026", "year": 2026, "code_url": "https://github.com/cvsp-lab/3DZip", "code_type": "Code", "contribution": "Combines coarse voxelization, DPP-based feature-diverse anchor selection, and spatially constrained merging for 3D VLM tokens", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Diversity-Guided" } ], "id": "3dzip-spatial-aware-feature-diversity-guided-token-compression-for-3d-question-answering", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-10-20", "added_in": "162519ad40579f9814e2a0f52443728ea11d6298", "bibtex": "@misc{3dzipspatialawarefeaturediversityguidedtok2026,\n title = {3DZip: Spatial-Aware Feature Diversity-Guided Token Compression for 3D Question Answering},\n year = {2026},\n howpublished = {ECCV 2026},\n url = {https://arxiv.org/abs/2608.01185}\n}" }, { "title": "AgilePruner: An Empirical Study of Attention and Diversity for Adaptive Visual Token Pruning in Large Vision-Language Models", "paper_url": "https://arxiv.org/abs/2603.01236", "venue": "ICLR 2026", "year": 2026, "code_url": "https://github.com/cvsp-lab/AgilePruner", "code_type": "Code", "contribution": "Combines CLS-attention selection with erank-based diversity pruning, adapting the balance to image complexity", "placements": [ { "stage": "Encoding", "category": "Encoding-Oriented Token Compression", "subcategory": "Attention-Aware" } ], "id": "agilepruner-an-empirical-study-of-attention-and-diversity-for-adaptive-visual-token-pruning-in-large-vision-language-models", "stages": [ "Encoding" ], "categories": [ "Encoding-Oriented Token Compression" ], "added_on": "2026-10-20", "added_in": "162519ad40579f9814e2a0f52443728ea11d6298", "bibtex": "@misc{agilepruneranempiricalstudyofattentionandd2026,\n title = {AgilePruner: An Empirical Study of Attention and Diversity for Adaptive Visual Token Pruning in Large Vision-Language Models},\n year = {2026},\n howpublished = {ICLR 2026},\n url = {https://arxiv.org/abs/2603.01236}\n}" }, { "title": "AVTP: Multi-Image Visual Token Pruning in Large Visual Language Models", "paper_url": "https://arxiv.org/abs/2608.26806", "venue": "EMNLP Findings 2026", "year": 2026, "code_url": "https://github.com/zry13/AVTP", "code_type": "Code", "contribution": "Adaptively assigns token budgets across images and supports diverse VLM architectures", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Attention-Guided" } ], "id": "avtp-multi-image-visual-token-pruning-in-large-visual-language-models", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-09-21", "added_in": "fd9fa236feead15f9fc0e8a157c3b78ae56f471d", "bibtex": "@misc{avtpmultiimagevisualtokenpruninginlargevis2026,\n title = {AVTP: Multi-Image Visual Token Pruning in Large Visual Language Models},\n year = {2026},\n howpublished = {EMNLP Findings 2026},\n url = {https://arxiv.org/abs/2608.26806}\n}" }, { "title": "CoverPruner: Who Speaks for the Pruned? Visual Token Pruning as Coverage Optimization", "paper_url": "https://arxiv.org/abs/2609.03158", "venue": "arXiv 2026", "year": 2026, "code_url": null, "code_type": null, "contribution": "Formulates pruning as query-weighted representational coverage maximization", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Diversity-Guided" } ], "id": "coverpruner-who-speaks-for-the-pruned-visual-token-pruning-as-coverage-optimization", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-09-21", "added_in": "fd9fa236feead15f9fc0e8a157c3b78ae56f471d", "bibtex": "@misc{coverprunerwhospeaksfortheprunedvisualtoke2026,\n title = {CoverPruner: Who Speaks for the Pruned? Visual Token Pruning as Coverage Optimization},\n year = {2026},\n howpublished = {arXiv 2026},\n url = {https://arxiv.org/abs/2609.03158}\n}" }, { "title": "DSTP: Why and When Visual Token Pruning Fails?", "paper_url": "https://arxiv.org/abs/2604.12358", "venue": "arXiv 2026", "year": 2026, "code_url": null, "code_type": null, "contribution": "Tracks decoding-stage relevant visual information shift to adapt token pruning during complex reasoning", "placements": [ { "stage": "Decoding", "category": "KV Cache Compression", "subcategory": "Token-Level" } ], "id": "dstp-why-and-when-visual-token-pruning-fails", "stages": [ "Decoding" ], "categories": [ "KV Cache Compression" ], "added_on": "2026-09-21", "added_in": "fd9fa236feead15f9fc0e8a157c3b78ae56f471d", "bibtex": "@misc{dstpwhyandwhenvisualtokenpruningfails2026,\n title = {DSTP: Why and When Visual Token Pruning Fails?},\n year = {2026},\n howpublished = {arXiv 2026},\n url = {https://arxiv.org/abs/2604.12358}\n}" }, { "title": "EvoComp: Learning Visual Token Compression via Semantic-Guided Evolutionary Labeling", "paper_url": "https://arxiv.org/abs/2604.17087", "venue": "CVPR 2026", "year": 2026, "code_url": null, "code_type": null, "contribution": "Learns a lightweight visual token compressor from semantic-guided evolutionary token labels", "placements": [ { "stage": "Encoding", "category": "Encoding-Oriented Token Compression", "subcategory": "Attention-Free" } ], "id": "evocomp-learning-visual-token-compression-via-semantic-guided-evolutionary-labeling", "stages": [ "Encoding" ], "categories": [ "Encoding-Oriented Token Compression" ], "added_on": "2026-09-21", "added_in": "fd9fa236feead15f9fc0e8a157c3b78ae56f471d", "bibtex": "@misc{evocomplearningvisualtokencompressionviase2026,\n title = {EvoComp: Learning Visual Token Compression via Semantic-Guided Evolutionary Labeling},\n year = {2026},\n howpublished = {CVPR 2026},\n url = {https://arxiv.org/abs/2604.17087}\n}" }, { "title": "Hi-Lo Prune: Look at What You'll Lose before Pruning with Hierarchical Token Selection", "paper_url": "https://openaccess.thecvf.com/content/CVPR2026/html/Sun_Hi-Lo_Prune_Look_at_What_Youll_Lose_before_Pruning_with_CVPR_2026_paper.html", "venue": "CVPR 2026", "year": 2026, "code_url": "https://github.com/sealost/Hi-Lo_Prune", "code_type": "Code", "contribution": "Hierarchically selects pruning candidates and transfers their information to retained tokens before early-layer removal", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Attention-Guided" } ], "id": "hi-lo-prune-look-at-what-you-ll-lose-before-pruning-with-hierarchical-token-selection", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-10-20", "added_in": "bf8daaed7993c2258fb8dc983ed32c4964e94ce9", "bibtex": "@misc{hiloprunelookatwhatyoulllosebeforepruningw2026,\n title = {Hi-Lo Prune: Look at What You'll Lose before Pruning with Hierarchical Token Selection},\n year = {2026},\n howpublished = {CVPR 2026},\n url = {https://openaccess.thecvf.com/content/CVPR2026/html/Sun_Hi-Lo_Prune_Look_at_What_Youll_Lose_before_Pruning_with_CVPR_2026_paper.html}\n}" }, { "title": "HybridKV: Hybrid KV Cache Compression for Efficient Multimodal Large Language Model Inference", "paper_url": "https://arxiv.org/abs/2604.05887", "venue": "ACL 2026", "year": 2026, "code_url": null, "code_type": null, "contribution": "Classifies attention heads as static or dynamic and applies head-specific pruning or chunk retrieval", "placements": [ { "stage": "Decoding", "category": "KV Cache Compression", "subcategory": "Head-Level" } ], "id": "hybridkv-hybrid-kv-cache-compression-for-efficient-multimodal-large-language-model-inference", "stages": [ "Decoding" ], "categories": [ "KV Cache Compression" ], "added_on": "2026-09-21", "added_in": "fd9fa236feead15f9fc0e8a157c3b78ae56f471d", "bibtex": "@misc{hybridkvhybridkvcachecompressionforefficie2026,\n title = {HybridKV: Hybrid KV Cache Compression for Efficient Multimodal Large Language Model Inference},\n year = {2026},\n howpublished = {ACL 2026},\n url = {https://arxiv.org/abs/2604.05887}\n}" }, { "title": "ICCTP: Instruction-Guided Cross-Modal Clustering for Training-Free Visual Token Pruning", "paper_url": "https://doi.org/10.1609/aaai.v40i14.38212", "venue": "AAAI 2026", "year": 2026, "code_url": null, "code_type": null, "contribution": "Uses instruction noun anchors for cross-modal clustering and high-ratio token pruning", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Attention-Guided" } ], "id": "icctp-instruction-guided-cross-modal-clustering-for-training-free-visual-token-pruning", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-09-21", "added_in": "fd9fa236feead15f9fc0e8a157c3b78ae56f471d", "bibtex": "@misc{icctpinstructionguidedcrossmodalclustering2026,\n title = {ICCTP: Instruction-Guided Cross-Modal Clustering for Training-Free Visual Token Pruning},\n year = {2026},\n howpublished = {AAAI 2026},\n url = {https://doi.org/10.1609/aaai.v40i14.38212}\n}" }, { "title": "Jag: Direct Box Prediction for Efficient Visual Grounding", "paper_url": "https://arxiv.org/abs/2610.13239", "venue": "arXiv 2026", "year": 2026, "code_url": "https://github.com/xyzzzh/Jag", "code_type": "Code", "contribution": "Adapts a compact multimodal model to predict a continuous bounding box from an existing input-token state in one forward pass, avoiding autoregressive coordinate generation.", "placements": [ { "stage": "Decoding", "category": "Direct Prediction", "subcategory": null } ], "id": "jag-direct-box-prediction-for-efficient-visual-grounding", "stages": [ "Decoding" ], "categories": [ "Direct Prediction" ], "added_on": "2026-10-20", "added_in": "b91d7270a30bcc4659bbbfb7e9c30688af5b570a", "bibtex": "@misc{jagdirectboxpredictionforefficientvisualgr2026,\n title = {Jag: Direct Box Prediction for Efficient Visual Grounding},\n year = {2026},\n howpublished = {arXiv 2026},\n url = {https://arxiv.org/abs/2610.13239}\n}" }, { "title": "MMSpec: Benchmarking Speculative Decoding for Vision-Language Models", "paper_url": "https://arxiv.org/abs/2603.14989", "venue": "arXiv 2026", "year": 2026, "code_url": "https://mmspec-bench.github.io/", "code_type": "Project page", "contribution": "Benchmarks ten speculative decoding methods and introduces vision-adaptive ViSkip", "placements": [ { "stage": "Decoding", "category": "Speculative Decoding", "subcategory": "Training-Free" } ], "id": "mmspec-benchmarking-speculative-decoding-for-vision-language-models", "stages": [ "Decoding" ], "categories": [ "Speculative Decoding" ], "added_on": "2026-09-21", "added_in": "fd9fa236feead15f9fc0e8a157c3b78ae56f471d", "bibtex": "@misc{mmspecbenchmarkingspeculativedecodingforvi2026,\n title = {MMSpec: Benchmarking Speculative Decoding for Vision-Language Models},\n year = {2026},\n howpublished = {arXiv 2026},\n url = {https://arxiv.org/abs/2603.14989}\n}" }, { "title": "OccamToken: Efficient VLM Inference with Training-Free and Budget-Adaptive Token Pruning", "paper_url": "https://arxiv.org/abs/2605.29657", "venue": "arXiv 2026", "year": 2026, "code_url": null, "code_type": null, "contribution": "Uses register-anchored relative evidence tests for image- and query-adaptive pruning", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Attention-Guided" } ], "id": "occamtoken-efficient-vlm-inference-with-training-free-and-budget-adaptive-token-pruning", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-09-21", "added_in": "fd9fa236feead15f9fc0e8a157c3b78ae56f471d", "bibtex": "@misc{occamtokenefficientvlminferencewithtrainin2026,\n title = {OccamToken: Efficient VLM Inference with Training-Free and Budget-Adaptive Token Pruning},\n year = {2026},\n howpublished = {arXiv 2026},\n url = {https://arxiv.org/abs/2605.29657}\n}" }, { "title": "Q Cache: Visual Attention is Valuable in Less than Half of Decode Layers for Multimodal Large Language Model", "paper_url": "https://arxiv.org/abs/2602.01901", "venue": "arXiv 2026", "year": 2026, "code_url": null, "code_type": null, "contribution": "Reuses similar attention queries across adjacent layers through a lightweight layer-shared cache", "placements": [ { "stage": "Decoding", "category": "Attention Reuse", "subcategory": null } ], "id": "q-cache-visual-attention-is-valuable-in-less-than-half-of-decode-layers-for-multimodal-large-language-model", "stages": [ "Decoding" ], "categories": [ "Attention Reuse" ], "added_on": "2026-09-21", "added_in": "fd9fa236feead15f9fc0e8a157c3b78ae56f471d", "bibtex": "@misc{qcachevisualattentionisvaluableinlessthanh2026,\n title = {Q Cache: Visual Attention is Valuable in Less than Half of Decode Layers for Multimodal Large Language Model},\n year = {2026},\n howpublished = {arXiv 2026},\n url = {https://arxiv.org/abs/2602.01901}\n}" }, { "title": "Semantic-Guided Slow-Fast Pruning of Visual Tokens for Vision-Language Models", "paper_url": "https://ieeexplore.ieee.org/abstract/document/11460400", "venue": "ICASSP 2026", "year": 2026, "code_url": null, "code_type": null, "contribution": "Semantic-guided slow-fast pruning of visual tokens for VLMs", "placements": [ { "stage": "Encoding", "category": "Encoding-Oriented Token Compression", "subcategory": "Attention-Aware" } ], "id": "semantic-guided-slow-fast-pruning-of-visual-tokens-for-vision-language-models", "stages": [ "Encoding" ], "categories": [ "Encoding-Oriented Token Compression" ], "added_on": "2026-09-21", "added_in": "83bb2d56348b3c0ba8a8c18c435389561cac1c10", "bibtex": "@misc{semanticguidedslowfastpruningofvisualtoken2026,\n title = {Semantic-Guided Slow-Fast Pruning of Visual Tokens for Vision-Language Models},\n year = {2026},\n howpublished = {ICASSP 2026},\n url = {https://ieeexplore.ieee.org/abstract/document/11460400}\n}" }, { "title": "SIEVE: When Vision Becomes Text", "paper_url": "https://arxiv.org/abs/2608.10489", "venue": "arXiv 2026", "year": 2026, "code_url": null, "code_type": null, "contribution": "Retains visual information unexplained by the text subspace using cross-modal residual guidance", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Attention-Guided" } ], "id": "sieve-when-vision-becomes-text", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-09-21", "added_in": "fd9fa236feead15f9fc0e8a157c3b78ae56f471d", "bibtex": "@misc{sievewhenvisionbecomestext2026,\n title = {SIEVE: When Vision Becomes Text},\n year = {2026},\n howpublished = {arXiv 2026},\n url = {https://arxiv.org/abs/2608.10489}\n}" }, { "title": "STC: Accelerating Streaming Video Large Language Models via Hierarchical Token Compression", "paper_url": "https://arxiv.org/abs/2512.00891", "venue": "CVPR 2026", "year": 2026, "code_url": "https://github.com/lern-to-write/STC", "code_type": "Code", "contribution": "Jointly accelerates ViT encoding through cross-frame feature caching and LLM prefilling through spatiotemporal token pruning", "placements": [ { "stage": "Encoding", "category": "Efficient Vision Encoders", "subcategory": "Video-Related" } ], "id": "stc-accelerating-streaming-video-large-language-models-via-hierarchical-token-compression", "stages": [ "Encoding" ], "categories": [ "Efficient Vision Encoders" ], "added_on": "2026-09-21", "added_in": "fd9fa236feead15f9fc0e8a157c3b78ae56f471d", "bibtex": "@misc{stcacceleratingstreamingvideolargelanguage2026,\n title = {STC: Accelerating Streaming Video Large Language Models via Hierarchical Token Compression},\n year = {2026},\n howpublished = {CVPR 2026},\n url = {https://arxiv.org/abs/2512.00891}\n}" }, { "title": "SwiftVLM: Efficient Vision-Language Model Inference via Cross-Layer Token Bypass", "paper_url": "https://arxiv.org/abs/2602.03134", "venue": "arXiv 2026", "year": 2026, "code_url": null, "code_type": null, "contribution": "Bypasses rather than permanently discards unselected tokens, enabling later-layer re-evaluation", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Attention-Guided" } ], "id": "swiftvlm-efficient-vision-language-model-inference-via-cross-layer-token-bypass", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-09-21", "added_in": "fd9fa236feead15f9fc0e8a157c3b78ae56f471d", "bibtex": "@misc{swiftvlmefficientvisionlanguagemodelinfere2026,\n title = {SwiftVLM: Efficient Vision-Language Model Inference via Cross-Layer Token Bypass},\n year = {2026},\n howpublished = {arXiv 2026},\n url = {https://arxiv.org/abs/2602.03134}\n}" }, { "title": "Unified Spatiotemporal Token Compression for Video-LLMs at Ultra-Low Retention", "paper_url": "https://arxiv.org/abs/2603.21957", "venue": "CVPR 2026", "year": 2026, "code_url": null, "code_type": null, "contribution": "Globally allocates a unified spatiotemporal token budget using contribution and redundancy signals", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Diversity-Guided" } ], "id": "unified-spatiotemporal-token-compression-for-video-llms-at-ultra-low-retention", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-09-21", "added_in": "fd9fa236feead15f9fc0e8a157c3b78ae56f471d", "bibtex": "@misc{unifiedspatiotemporaltokencompressionforvi2026,\n title = {Unified Spatiotemporal Token Compression for Video-LLMs at Ultra-Low Retention},\n year = {2026},\n howpublished = {CVPR 2026},\n url = {https://arxiv.org/abs/2603.21957}\n}" }, { "title": "Adaptive Fast-and-Slow Visual Program Reasoning for Long-Form VideoQA", "paper_url": "https://arxiv.org/abs/2509.17743", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Fast-slow reasoning framework routing simple queries to VideoLLM and complex ones to visual program workflow", "placements": [ { "stage": "Decoding", "category": "Efficient Reasoning", "subcategory": null } ], "id": "adaptive-fast-and-slow-visual-program-reasoning-for-long-form-videoqa", "stages": [ "Decoding" ], "categories": [ "Efficient Reasoning" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{adaptivefastandslowvisualprogramreasoningf2025,\n title = {Adaptive Fast-and-Slow Visual Program Reasoning for Long-Form VideoQA},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2509.17743}\n}" }, { "title": "AirCache: Activating Inter-modal Relevancy KV Cache Compression", "paper_url": "https://arxiv.org/abs/2503.23956", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Models stable inter-modal relevance and adaptively allocates visual KV budgets across layers", "placements": [ { "stage": "Decoding", "category": "KV Cache Compression", "subcategory": "Layer-Level" } ], "id": "aircache-activating-inter-modal-relevancy-kv-cache-compression", "stages": [ "Decoding" ], "categories": [ "KV Cache Compression" ], "added_on": "2026-09-21", "added_in": "fd9fa236feead15f9fc0e8a157c3b78ae56f471d", "bibtex": "@misc{aircacheactivatingintermodalrelevancykvcac2025,\n title = {AirCache: Activating Inter-modal Relevancy KV Cache Compression},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2503.23956}\n}" }, { "title": "AKVQ-VL: Attention-Aware KV Cache Adaptive 2-Bit Quantization", "paper_url": "https://arxiv.org/abs/2501.15021", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Adaptive mixed-precision quantization with high bit-width for critical tokens and 2-bit for others", "placements": [ { "stage": "Decoding", "category": "KV Cache Compression", "subcategory": "Bit-Level" } ], "id": "akvq-vl-attention-aware-kv-cache-adaptive-2-bit-quantization", "stages": [ "Decoding" ], "categories": [ "KV Cache Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{akvqvlattentionawarekvcacheadaptive2bitqua2025,\n title = {AKVQ-VL: Attention-Aware KV Cache Adaptive 2-Bit Quantization},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2501.15021}\n}" }, { "title": "ATP-LLaVA: Adaptive token pruning for LVLMs", "paper_url": "https://openaccess.thecvf.com/content/CVPR2025/html/Ye_ATP-LLaVA_Adaptive_Token_Pruning_for_Large_Vision_Language_Models_CVPR_2025_paper.html", "venue": "CVPR 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Learnable adaptive token pruning module computing importance score", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Attention-Guided" } ], "id": "atp-llava-adaptive-token-pruning-for-lvlms", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{atpllavaadaptivetokenpruningforlvlms2025,\n title = {ATP-LLaVA: Adaptive token pruning for LVLMs},\n year = {2025},\n howpublished = {CVPR 2025},\n url = {https://openaccess.thecvf.com/content/CVPR2025/html/Ye_ATP-LLaVA_Adaptive_Token_Pruning_for_Large_Vision_Language_Models_CVPR_2025_paper.html}\n}" }, { "title": "Beyond Text-Visual Attention: Exploiting Visual Cues for Effective Token Pruning", "paper_url": "https://arxiv.org/abs/2412.01818", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Selects informative tokens using visual attention (VisPruner)", "placements": [ { "stage": "Encoding", "category": "Encoding-Oriented Token Compression", "subcategory": "Attention-Free" } ], "id": "beyond-text-visual-attention-exploiting-visual-cues-for-effective-token-pruning", "stages": [ "Encoding" ], "categories": [ "Encoding-Oriented Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{beyondtextvisualattentionexploitingvisualc2025,\n title = {Beyond Text-Visual Attention: Exploiting Visual Cues for Effective Token Pruning},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2412.01818}\n}" }, { "title": "BTP: Balanced Token Pruning", "paper_url": "https://arxiv.org/abs/2505.22038", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Multi-stage pruning with diversity and attention ranking objectives", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Attention-Guided" } ], "id": "btp-balanced-token-pruning", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{btpbalancedtokenpruning2025,\n title = {BTP: Balanced Token Pruning},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2505.22038}\n}" }, { "title": "CalibQuant: 1-Bit KV Cache Quantization for Multimodal LLMs", "paper_url": "https://arxiv.org/abs/2502.14882", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Channel-wise 1-bit quantization with post-calibration for extreme values", "placements": [ { "stage": "Decoding", "category": "KV Cache Compression", "subcategory": "Bit-Level" } ], "id": "calibquant-1-bit-kv-cache-quantization-for-multimodal-llms", "stages": [ "Decoding" ], "categories": [ "KV Cache Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{calibquant1bitkvcachequantizationformultim2025,\n title = {CalibQuant: 1-Bit KV Cache Quantization for Multimodal LLMs},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2502.14882}\n}" }, { "title": "CDPruner: Maximizing Conditional Diversity for Token Pruning", "paper_url": "https://arxiv.org/abs/2506.10967", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Determinantal Point Processes (DPP) maximizing conditional diversity", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Diversity-Guided" } ], "id": "cdpruner-maximizing-conditional-diversity-for-token-pruning", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{cdprunermaximizingconditionaldiversityfort2025,\n title = {CDPruner: Maximizing Conditional Diversity for Token Pruning},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2506.10967}\n}" }, { "title": "DART: Stop looking for important tokens, duplication matters more", "paper_url": "https://arxiv.org/abs/2502.11494", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Pivot-based duplication pruning selecting tokens with low duplication", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Diversity-Guided" } ], "id": "dart-stop-looking-for-important-tokens-duplication-matters-more", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{dartstoplookingforimportanttokensduplicati2025,\n title = {DART: Stop looking for important tokens, duplication matters more},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2502.11494}\n}" }, { "title": "DivPrune: Diversity-based visual token pruning", "paper_url": "https://openaccess.thecvf.com/content/CVPR2025/html/Alvar_DivPrune_Diversity-based_Visual_Token_Pruning_for_Large_Multimodal_Models_CVPR_2025_paper.html", "venue": "CVPR 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Max-Min diversity optimization for token subset selection", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Diversity-Guided" } ], "id": "divprune-diversity-based-visual-token-pruning", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{divprunediversitybasedvisualtokenpruning2025,\n title = {DivPrune: Diversity-based visual token pruning},\n year = {2025},\n howpublished = {CVPR 2025},\n url = {https://openaccess.thecvf.com/content/CVPR2025/html/Alvar_DivPrune_Diversity-based_Visual_Token_Pruning_for_Large_Multimodal_Models_CVPR_2025_paper.html}\n}" }, { "title": "DyCoke: Dynamic Compression of Tokens for Fast Video LLMs", "paper_url": "https://openaccess.thecvf.com/content/CVPR2025/html/Tao_DyCoke_Dynamic_Compression_of_Tokens_for_Fast_Video_Large_Language_CVPR_2025_paper.html", "venue": "CVPR 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Plug-and-play temporal compression module minimizing temporal redundancy", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Diversity-Guided" } ], "id": "dycoke-dynamic-compression-of-tokens-for-fast-video-llms", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{dycokedynamiccompressionoftokensforfastvid2025,\n title = {DyCoke: Dynamic Compression of Tokens for Fast Video LLMs},\n year = {2025},\n howpublished = {CVPR 2025},\n url = {https://openaccess.thecvf.com/content/CVPR2025/html/Tao_DyCoke_Dynamic_Compression_of_Tokens_for_Fast_Video_Large_Language_CVPR_2025_paper.html}\n}" }, { "title": "FastCache: Optimizing multimodal LLM serving", "paper_url": "https://arxiv.org/abs/2503.08461", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Lightweight modality-specific compressor learning compression patterns", "placements": [ { "stage": "Decoding", "category": "KV Cache Compression", "subcategory": "Token-Level" } ], "id": "fastcache-optimizing-multimodal-llm-serving", "stages": [ "Decoding" ], "categories": [ "KV Cache Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{fastcacheoptimizingmultimodalllmserving2025,\n title = {FastCache: Optimizing multimodal LLM serving},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2503.08461}\n}" }, { "title": "FastVID: Dynamic density pruning for fast video LLMs", "paper_url": "https://arxiv.org/abs/2503.11187", "venue": "arXiv 2025", "year": 2025, "code_url": "https://github.com/LunarShen/FastVID", "code_type": "Code", "contribution": "Temporal segmentation and density spatiotemporal pruning", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Diversity-Guided" } ], "id": "fastvid-dynamic-density-pruning-for-fast-video-llms", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{fastviddynamicdensitypruningforfastvideoll2025,\n title = {FastVID: Dynamic density pruning for fast video LLMs},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2503.11187}\n}" }, { "title": "FastVLM: Efficient vision encoding for vision language models", "paper_url": "https://openaccess.thecvf.com/content/CVPR2025/html/Vasu_FastVLM_Efficient_Vision_Encoding_for_Vision_Language_Models_CVPR_2025_paper.html", "venue": "CVPR 2025", "year": 2025, "code_url": "https://machinelearning.apple.com/research/fastvlm-efficient-vision-encoding", "code_type": "Project page", "contribution": "Hybrid vision encoder outputting fewer tokens and reducing encoding time", "placements": [ { "stage": "Encoding", "category": "Efficient Vision Encoders", "subcategory": "Image-Related" } ], "id": "fastvlm-efficient-vision-encoding-for-vision-language-models", "stages": [ "Encoding" ], "categories": [ "Efficient Vision Encoders" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{fastvlmefficientvisionencodingforvisionlan2025,\n title = {FastVLM: Efficient vision encoding for vision language models},\n year = {2025},\n howpublished = {CVPR 2025},\n url = {https://openaccess.thecvf.com/content/CVPR2025/html/Vasu_FastVLM_Efficient_Vision_Encoding_for_Vision_Language_Models_CVPR_2025_paper.html}\n}" }, { "title": "FastVLM: Self-Speculative Decoding for Fast Vision-Language Model Inference", "paper_url": "https://arxiv.org/abs/2510.22641", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Imitation-based draft model learning from deeper representations (self-speculative decoding)", "placements": [ { "stage": "Decoding", "category": "Speculative Decoding", "subcategory": "Training-Aware" } ], "id": "fastvlm-self-speculative-decoding-for-fast-vision-language-model-inference", "stages": [ "Decoding" ], "categories": [ "Speculative Decoding" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{fastvlmselfspeculativedecodingforfastvisio2025,\n title = {FastVLM: Self-Speculative Decoding for Fast Vision-Language Model Inference},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2510.22641}\n}" }, { "title": "Fit and Prune: Fast and training-free visual token pruning", "paper_url": "https://doi.org/10.1609/aaai.v39i21.34366", "venue": "AAAI 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Minimizes divergence of attention distributions before and after pruning", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Attention-Guided" } ], "id": "fit-and-prune-fast-and-training-free-visual-token-pruning", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{fitandprunefastandtrainingfreevisualtokenp2025,\n title = {Fit and Prune: Fast and training-free visual token pruning},\n year = {2025},\n howpublished = {AAAI 2025},\n url = {https://doi.org/10.1609/aaai.v39i21.34366}\n}" }, { "title": "FLASH: Latent-Aware Semi-Autoregressive Speculative Decoding", "paper_url": "https://arxiv.org/abs/2505.12728", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Visual token compression mechanism and semi-autoregressive head for draft model optimization", "placements": [ { "stage": "Decoding", "category": "Speculative Decoding", "subcategory": "Training-Aware" } ], "id": "flash-latent-aware-semi-autoregressive-speculative-decoding", "stages": [ "Decoding" ], "categories": [ "Speculative Decoding" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{flashlatentawaresemiautoregressivespeculat2025,\n title = {FLASH: Latent-Aware Semi-Autoregressive Speculative Decoding},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2505.12728}\n}" }, { "title": "FOCUS: Efficient Keyframe Selection for Long Video Understanding", "paper_url": "https://arxiv.org/abs/2510.27280", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Formulates keyframe selection as combinatorial pure-exploration", "placements": [ { "stage": "Encoding", "category": "Keyframe Selection", "subcategory": "Training-Free" } ], "id": "focus-efficient-keyframe-selection-for-long-video-understanding", "stages": [ "Encoding" ], "categories": [ "Keyframe Selection" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{focusefficientkeyframeselectionforlongvide2025,\n title = {FOCUS: Efficient Keyframe Selection for Long Video Understanding},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2510.27280}\n}" }, { "title": "FOLDER: Accelerating Multi-modal Large Language Models", "paper_url": "https://arxiv.org/abs/2501.02430", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Plug-and-play module in final vision backbone blocks for merging operations", "placements": [ { "stage": "Encoding", "category": "Encoding-Oriented Token Compression", "subcategory": "Attention-Free" } ], "id": "folder-accelerating-multi-modal-large-language-models", "stages": [ "Encoding" ], "categories": [ "Encoding-Oriented Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{folderacceleratingmultimodallargelanguagem2025,\n title = {FOLDER: Accelerating Multi-modal Large Language Models},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2501.02430}\n}" }, { "title": "G-Prune: Training-free visual token pruning from graph perspective", "paper_url": "https://doi.org/10.1609/aaai.v39i4.32427", "venue": "AAAI 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Similarity graph and information flow to retain representative tokens", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Diversity-Guided" } ], "id": "g-prune-training-free-visual-token-pruning-from-graph-perspective", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{gprunetrainingfreevisualtokenpruningfromgr2025,\n title = {G-Prune: Training-free visual token pruning from graph perspective},\n year = {2025},\n howpublished = {AAAI 2025},\n url = {https://doi.org/10.1609/aaai.v39i4.32427}\n}" }, { "title": "Glyph: Scaling Context Windows via Visual-Text Compression", "paper_url": "https://arxiv.org/abs/2510.17800", "venue": "arXiv 2025", "year": 2025, "code_url": "https://github.com/thu-coai/Glyph", "code_type": "Code", "contribution": "DeepEncoder maintaining low activations under high-resolution input", "placements": [ { "stage": "Encoding", "category": "Efficient Vision Encoders", "subcategory": "Image-Related" }, { "stage": "Decoding", "category": "Speculative Decoding", "subcategory": "Training-Aware" } ], "id": "glyph-scaling-context-windows-via-visual-text-compression", "stages": [ "Decoding", "Encoding" ], "categories": [ "Efficient Vision Encoders", "Speculative Decoding" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{glyphscalingcontextwindowsviavisualtextcom2025,\n title = {Glyph: Scaling Context Windows via Visual-Text Compression},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2510.17800}\n}" }, { "title": "HIVTP: Hierarchical Visual Token Pruning", "paper_url": "https://arxiv.org/abs/2509.23663", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Attention maps from middle encoder layers to estimate visual token importance", "placements": [ { "stage": "Encoding", "category": "Encoding-Oriented Token Compression", "subcategory": "Attention-Aware" } ], "id": "hivtp-hierarchical-visual-token-pruning", "stages": [ "Encoding" ], "categories": [ "Encoding-Oriented Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{hivtphierarchicalvisualtokenpruning2025,\n title = {HIVTP: Hierarchical Visual Token Pruning},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2509.23663}\n}" }, { "title": "HoliTom: Holistic Token Merging for Fast Video LLMs", "paper_url": "https://arxiv.org/abs/2505.21334", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Global redundancy-aware segmentation followed by spatiotemporal merging", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Diversity-Guided" } ], "id": "holitom-holistic-token-merging-for-fast-video-llms", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{holitomholistictokenmergingforfastvideollm2025,\n title = {HoliTom: Holistic Token Merging for Fast Video LLMs},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2505.21334}\n}" }, { "title": "InfiniPot-V: Memory-Constrained KV Cache Compression", "paper_url": "https://arxiv.org/abs/2506.15745", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Layer-wise adaptive pooling with varying kernel sizes to balance abstraction and detail", "placements": [ { "stage": "Decoding", "category": "KV Cache Compression", "subcategory": "Layer-Level" } ], "id": "infinipot-v-memory-constrained-kv-cache-compression", "stages": [ "Decoding" ], "categories": [ "KV Cache Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{infinipotvmemoryconstrainedkvcachecompress2025,\n title = {InfiniPot-V: Memory-Constrained KV Cache Compression},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2506.15745}\n}" }, { "title": "LiveVLM: Efficient Online Video Understanding via Streaming-Oriented KV Cache", "paper_url": "https://arxiv.org/abs/2505.15269", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Dual-memory approach with short-term sliding window and compressed long-term memory", "placements": [ { "stage": "Decoding", "category": "KV Cache Compression", "subcategory": "Token-Level" } ], "id": "livevlm-efficient-online-video-understanding-via-streaming-oriented-kv-cache", "stages": [ "Decoding" ], "categories": [ "KV Cache Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{livevlmefficientonlinevideounderstandingvi2025,\n title = {LiveVLM: Efficient Online Video Understanding via Streaming-Oriented KV Cache},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2505.15269}\n}" }, { "title": "LLaVA-PruMerge: Adaptive Token Reduction", "paper_url": "https://arxiv.org/abs/2403.15388", "venue": "ICCV 2025", "year": 2025, "code_url": "https://github.com/42Shawn/LLaVA-PruMerge", "code_type": "Code", "contribution": "Reduces visual tokens according to similarities between class and spatial tokens", "placements": [ { "stage": "Encoding", "category": "Encoding-Oriented Token Compression", "subcategory": "Attention-Free" } ], "id": "llava-prumerge-adaptive-token-reduction", "stages": [ "Encoding" ], "categories": [ "Encoding-Oriented Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{llavaprumergeadaptivetokenreduction2025,\n title = {LLaVA-PruMerge: Adaptive Token Reduction},\n year = {2025},\n howpublished = {ICCV 2025},\n url = {https://arxiv.org/abs/2403.15388}\n}" }, { "title": "LongVU: Spatiotemporal Adaptive Compression", "paper_url": "https://openreview.net/forum?id=XzZC4gs1mf", "venue": "ICML 2025", "year": 2025, "code_url": "https://vision-cair.github.io/LongVU/", "code_type": "Project page", "contribution": "Preserves full features for query-relevant frames", "placements": [ { "stage": "Encoding", "category": "Adaptive Resolution", "subcategory": null } ], "id": "longvu-spatiotemporal-adaptive-compression", "stages": [ "Encoding" ], "categories": [ "Adaptive Resolution" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{longvuspatiotemporaladaptivecompression2025,\n title = {LongVU: Spatiotemporal Adaptive Compression},\n year = {2025},\n howpublished = {ICML 2025},\n url = {https://openreview.net/forum?id=XzZC4gs1mf}\n}" }, { "title": "LongVU: Spatiotemporal Adaptive Compression for Long Video-Language Understanding", "paper_url": "https://openreview.net/forum?id=XzZC4gs1mf", "venue": "ICML 2025", "year": 2025, "code_url": "https://vision-cair.github.io/LongVU/", "code_type": "Project page", "contribution": "Preserves full features for query-relevant frames while applying spatial pooling", "placements": [ { "stage": "Encoding", "category": "Efficient Vision Encoders", "subcategory": "Video-Related" } ], "id": "longvu-spatiotemporal-adaptive-compression-for-long-video-language-understanding", "stages": [ "Encoding" ], "categories": [ "Efficient Vision Encoders" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{longvuspatiotemporaladaptivecompressionfor2025,\n title = {LongVU: Spatiotemporal Adaptive Compression for Long Video-Language Understanding},\n year = {2025},\n howpublished = {ICML 2025},\n url = {https://openreview.net/forum?id=XzZC4gs1mf}\n}" }, { "title": "M-LLM based video frame selection for efficient video understanding", "paper_url": "https://openaccess.thecvf.com/content/CVPR2025/html/Hu_M-LLM_Based_Video_Frame_Selection_for_Efficient_Video_Understanding_CVPR_2025_paper.html", "venue": "CVPR 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Uses spatial and temporal signals as supervision to train frame selector", "placements": [ { "stage": "Encoding", "category": "Keyframe Selection", "subcategory": "Training-Aware" } ], "id": "m-llm-based-video-frame-selection-for-efficient-video-understanding", "stages": [ "Encoding" ], "categories": [ "Keyframe Selection" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{mllmbasedvideoframeselectionforefficientvi2025,\n title = {M-LLM based video frame selection for efficient video understanding},\n year = {2025},\n howpublished = {CVPR 2025},\n url = {https://openaccess.thecvf.com/content/CVPR2025/html/Hu_M-LLM_Based_Video_Frame_Selection_for_Efficient_Video_Understanding_CVPR_2025_paper.html}\n}" }, { "title": "MadaKV: Adaptive Modality-Perception KV Cache Eviction", "paper_url": "https://aclanthology.org/2025.acl-long.652/", "venue": "ACL 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Inter-layer compensation mechanism dynamically adjusting budgets", "placements": [ { "stage": "Decoding", "category": "KV Cache Compression", "subcategory": "Layer-Level" } ], "id": "madakv-adaptive-modality-perception-kv-cache-eviction", "stages": [ "Decoding" ], "categories": [ "KV Cache Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{madakvadaptivemodalityperceptionkvcacheevi2025,\n title = {MadaKV: Adaptive Modality-Perception KV Cache Eviction},\n year = {2025},\n howpublished = {ACL 2025},\n url = {https://aclanthology.org/2025.acl-long.652/}\n}" }, { "title": "MASSV: Multimodal Adaptation and Self-Data Distillation for Speculative Decoding of VLMs", "paper_url": "https://arxiv.org/abs/2505.10526", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Converts small language models into multimodal drafters using a vision projector and self-distillation", "placements": [ { "stage": "Decoding", "category": "Speculative Decoding", "subcategory": "Training-Aware" } ], "id": "massv-multimodal-adaptation-and-self-data-distillation-for-speculative-decoding-of-vlms", "stages": [ "Decoding" ], "categories": [ "Speculative Decoding" ], "added_on": "2026-09-21", "added_in": "fd9fa236feead15f9fc0e8a157c3b78ae56f471d", "bibtex": "@misc{massvmultimodaladaptationandselfdatadistil2025,\n title = {MASSV: Multimodal Adaptation and Self-Data Distillation for Speculative Decoding of VLMs},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2505.10526}\n}" }, { "title": "Meda: Dynamic KV cache allocation for efficient multimodal inference", "paper_url": "https://arxiv.org/abs/2502.17599", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Cross-modal attention entropy guiding cache allocation to layers with complex interactions", "placements": [ { "stage": "Decoding", "category": "KV Cache Compression", "subcategory": "Layer-Level" } ], "id": "meda-dynamic-kv-cache-allocation-for-efficient-multimodal-inference", "stages": [ "Decoding" ], "categories": [ "KV Cache Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{medadynamickvcacheallocationforefficientmu2025,\n title = {Meda: Dynamic KV cache allocation for efficient multimodal inference},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2502.17599}\n}" }, { "title": "MMInference: Modality-Aware Permutation Sparse Attention", "paper_url": "https://arxiv.org/abs/2504.16083", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Identifies three distinct attention patterns in LVLMs with modality-aware permutation", "placements": [ { "stage": "Prefilling", "category": "Sparse Attention", "subcategory": null } ], "id": "mminference-modality-aware-permutation-sparse-attention", "stages": [ "Prefilling" ], "categories": [ "Sparse Attention" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{mminferencemodalityawarepermutationsparsea2025,\n title = {MMInference: Modality-Aware Permutation Sparse Attention},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2504.16083}\n}" }, { "title": "MSD: Speculative Decoding Reimagined for Multimodal Large Language Models", "paper_url": "https://arxiv.org/abs/2505.14260", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Two-stage training enabling draft model to acquire language modeling and visual perception", "placements": [ { "stage": "Decoding", "category": "Speculative Decoding", "subcategory": "Training-Aware" } ], "id": "msd-speculative-decoding-reimagined-for-multimodal-large-language-models", "stages": [ "Decoding" ], "categories": [ "Speculative Decoding" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{msdspeculativedecodingreimaginedformultimo2025,\n title = {MSD: Speculative Decoding Reimagined for Multimodal Large Language Models},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2505.14260}\n}" }, { "title": "PixelThink: Towards Efficient Chain-of-Pixel Reasoning", "paper_url": "https://arxiv.org/abs/2505.23727", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Reinforcement learning to regulate reasoning chain length based on task difficulty and model confidence", "placements": [ { "stage": "Decoding", "category": "Efficient Reasoning", "subcategory": null } ], "id": "pixelthink-towards-efficient-chain-of-pixel-reasoning", "stages": [ "Decoding" ], "categories": [ "Efficient Reasoning" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{pixelthinktowardsefficientchainofpixelreas2025,\n title = {PixelThink: Towards Efficient Chain-of-Pixel Reasoning},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2505.23727}\n}" }, { "title": "Prolonged reasoning is not all you need: Certainty-based adaptive routing", "paper_url": "https://arxiv.org/abs/2505.15154", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Certainty-based routing triggering long thought chains only when initial answer exhibits high uncertainty", "placements": [ { "stage": "Decoding", "category": "Efficient Reasoning", "subcategory": null } ], "id": "prolonged-reasoning-is-not-all-you-need-certainty-based-adaptive-routing", "stages": [ "Decoding" ], "categories": [ "Efficient Reasoning" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{prolongedreasoningisnotallyouneedcertainty2025,\n title = {Prolonged reasoning is not all you need: Certainty-based adaptive routing},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2505.15154}\n}" }, { "title": "PruneVid: Visual token pruning for efficient video LLMs", "paper_url": "https://aclanthology.org/2025.findings-acl.1024/", "venue": "ACL 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Spatiotemporal token merging before LLMs", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Diversity-Guided" } ], "id": "prunevid-visual-token-pruning-for-efficient-video-llms", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{prunevidvisualtokenpruningforefficientvide2025,\n title = {PruneVid: Visual token pruning for efficient video LLMs},\n year = {2025},\n howpublished = {ACL 2025},\n url = {https://aclanthology.org/2025.findings-acl.1024/}\n}" }, { "title": "Q-Frame: Query-aware Frame Selection and Multi-Resolution Adaptation", "paper_url": "https://arxiv.org/abs/2506.22139", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Text-image matching network with Gumbel-Max trick", "placements": [ { "stage": "Encoding", "category": "Keyframe Selection", "subcategory": "Training-Free" }, { "stage": "Encoding", "category": "Adaptive Resolution", "subcategory": null } ], "id": "q-frame-query-aware-frame-selection-and-multi-resolution-adaptation", "stages": [ "Encoding" ], "categories": [ "Adaptive Resolution", "Keyframe Selection" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{qframequeryawareframeselectionandmultireso2025,\n title = {Q-Frame: Query-aware Frame Selection and Multi-Resolution Adaptation},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2506.22139}\n}" }, { "title": "ReKV: Streaming video QA with in-context video KV-cache retrieval", "paper_url": "https://arxiv.org/abs/2503.00540", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Retrieval-based framework offloading video chunks to external memory", "placements": [ { "stage": "Decoding", "category": "KV Cache Compression", "subcategory": "Token-Level" } ], "id": "rekv-streaming-video-qa-with-in-context-video-kv-cache-retrieval", "stages": [ "Decoding" ], "categories": [ "KV Cache Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{rekvstreamingvideoqawithincontextvideokvca2025,\n title = {ReKV: Streaming video QA with in-context video KV-cache retrieval},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2503.00540}\n}" }, { "title": "SpargeAttn: Accurate sparse attention accelerating any model inference", "paper_url": "https://arxiv.org/abs/2502.18137", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Two-stage online filter to skip unimportant regions in sparse attention", "placements": [ { "stage": "Prefilling", "category": "Sparse Attention", "subcategory": null } ], "id": "spargeattn-accurate-sparse-attention-accelerating-any-model-inference", "stages": [ "Prefilling" ], "categories": [ "Sparse Attention" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{spargeattnaccuratesparseattentionaccelerat2025,\n title = {SpargeAttn: Accurate sparse attention accelerating any model inference},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2502.18137}\n}" }, { "title": "SparseMM: Head Sparsity Emerges from Visual Concept Responses", "paper_url": "https://arxiv.org/abs/2506.05344", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Identifies vital visual heads and allocates asymmetric budgets based on visual relevance", "placements": [ { "stage": "Decoding", "category": "KV Cache Compression", "subcategory": "Head-Level" } ], "id": "sparsemm-head-sparsity-emerges-from-visual-concept-responses", "stages": [ "Decoding" ], "categories": [ "KV Cache Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{sparsemmheadsparsityemergesfromvisualconce2025,\n title = {SparseMM: Head Sparsity Emerges from Visual Concept Responses},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2506.05344}\n}" }, { "title": "SparseVILA: Decoupling Visual Sparsity for Efficient VLM Inference", "paper_url": "https://arxiv.org/abs/2510.17777", "venue": "ICCV 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Estimates token importance from visual encoder's self-attention maps", "placements": [ { "stage": "Encoding", "category": "Encoding-Oriented Token Compression", "subcategory": "Attention-Aware" } ], "id": "sparsevila-decoupling-visual-sparsity-for-efficient-vlm-inference", "stages": [ "Encoding" ], "categories": [ "Encoding-Oriented Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{sparseviladecouplingvisualsparsityforeffic2025,\n title = {SparseVILA: Decoupling Visual Sparsity for Efficient VLM Inference},\n year = {2025},\n howpublished = {ICCV 2025},\n url = {https://arxiv.org/abs/2510.17777}\n}" }, { "title": "Spec-LLaVA: Accelerating VLMs with Dynamic Tree-Based Speculative Decoding", "paper_url": "https://arxiv.org/abs/2509.11961", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Compact distilled draft model paired with tree-based verification algorithm", "placements": [ { "stage": "Decoding", "category": "Speculative Decoding", "subcategory": "Training-Aware" } ], "id": "spec-llava-accelerating-vlms-with-dynamic-tree-based-speculative-decoding", "stages": [ "Decoding" ], "categories": [ "Speculative Decoding" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{specllavaacceleratingvlmswithdynamictreeba2025,\n title = {Spec-LLaVA: Accelerating VLMs with Dynamic Tree-Based Speculative Decoding},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2509.11961}\n}" }, { "title": "SpecVLM: Enhancing speculative decoding via verifier-guided token pruning", "paper_url": "https://doi.org/10.48550/arXiv.2508.16201", "venue": "EMNLP 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Verifier-guided staged pruning removing up to 90% of vision tokens from draft model input", "placements": [ { "stage": "Decoding", "category": "Speculative Decoding", "subcategory": "Training-Free" } ], "id": "specvlm-enhancing-speculative-decoding-via-verifier-guided-token-pruning", "stages": [ "Decoding" ], "categories": [ "Speculative Decoding" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{specvlmenhancingspeculativedecodingviaveri2025,\n title = {SpecVLM: Enhancing speculative decoding via verifier-guided token pruning},\n year = {2025},\n howpublished = {EMNLP 2025},\n url = {https://doi.org/10.48550/arXiv.2508.16201}\n}" }, { "title": "SpecVLM: Fast Speculative Decoding in Vision-Language Models", "paper_url": "https://arxiv.org/abs/2509.11815", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Elastic visual compressor adaptively selecting from multiple compression primitives", "placements": [ { "stage": "Decoding", "category": "Speculative Decoding", "subcategory": "Training-Aware" } ], "id": "specvlm-fast-speculative-decoding-in-vision-language-models", "stages": [ "Decoding" ], "categories": [ "Speculative Decoding" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{specvlmfastspeculativedecodinginvisionlang2025,\n title = {SpecVLM: Fast Speculative Decoding in Vision-Language Models},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2509.11815}\n}" }, { "title": "ST3: Accelerating MLLM by spatial-temporal visual token trimming", "paper_url": "https://doi.org/10.1609/aaai.v39i10.33201", "venue": "AAAI 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Progressive pruning of visual tokens in deeper layers based on decreasing visual importance", "placements": [ { "stage": "Decoding", "category": "KV Cache Compression", "subcategory": "Layer-Level" } ], "id": "st3-accelerating-mllm-by-spatial-temporal-visual-token-trimming", "stages": [ "Decoding" ], "categories": [ "KV Cache Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{st3acceleratingmllmbyspatialtemporalvisual2025,\n title = {ST3: Accelerating MLLM by spatial-temporal visual token trimming},\n year = {2025},\n howpublished = {AAAI 2025},\n url = {https://doi.org/10.1609/aaai.v39i10.33201}\n}" }, { "title": "StreamingTOM: Streaming Token Compression", "paper_url": "https://arxiv.org/abs/2510.18269", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Causal temporal reduction with fixed per-frame budget", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Diversity-Guided" } ], "id": "streamingtom-streaming-token-compression", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{streamingtomstreamingtokencompression2025,\n title = {StreamingTOM: Streaming Token Compression},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2510.18269}\n}" }, { "title": "StreamingVLM: Real-time understanding for infinite video streams", "paper_url": "https://arxiv.org/abs/2510.09608", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Maintains compact token subset by reusing attention sinks and recent token windows", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Attention-Guided" } ], "id": "streamingvlm-real-time-understanding-for-infinite-video-streams", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{streamingvlmrealtimeunderstandingforinfini2025,\n title = {StreamingVLM: Real-time understanding for infinite video streams},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2510.09608}\n}" }, { "title": "STTM: Multi-granular spatio-temporal token merging", "paper_url": "https://doi.org/10.48550/arXiv.2507.07990", "venue": "ICCV 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Quadtree spatial transformation with directed pairwise merging", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Diversity-Guided" } ], "id": "sttm-multi-granular-spatio-temporal-token-merging", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{sttmmultigranularspatiotemporaltokenmergin2025,\n title = {STTM: Multi-granular spatio-temporal token merging},\n year = {2025},\n howpublished = {ICCV 2025},\n url = {https://doi.org/10.48550/arXiv.2507.07990}\n}" }, { "title": "TimeChat-Online: 80% Visual Tokens are Naturally Redundant", "paper_url": "https://doi.org/10.48550/ARXIV.2504.17343", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Differential token drop module filtering redundant content in streaming videos", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Diversity-Guided" } ], "id": "timechat-online-80-visual-tokens-are-naturally-redundant", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{timechatonline80visualtokensarenaturallyre2025,\n title = {TimeChat-Online: 80\\% Visual Tokens are Naturally Redundant},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://doi.org/10.48550/ARXIV.2504.17343}\n}" }, { "title": "TokenPacker: Efficient Visual Projector for Multimodal LLM", "paper_url": "https://doi.org/10.1007/S11263-025-02491-7", "venue": "IJCV 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Coarse-to-fine scheme injecting enriched characteristics", "placements": [ { "stage": "Encoding", "category": "Efficient Modality Adapters", "subcategory": null } ], "id": "tokenpacker-efficient-visual-projector-for-multimodal-llm", "stages": [ "Encoding" ], "categories": [ "Efficient Modality Adapters" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{tokenpackerefficientvisualprojectorformult2025,\n title = {TokenPacker: Efficient Visual Projector for Multimodal LLM},\n year = {2025},\n howpublished = {IJCV 2025},\n url = {https://doi.org/10.1007/S11263-025-02491-7}\n}" }, { "title": "ToSA: Token Merging with Spatial Awareness", "paper_url": "https://arxiv.org/abs/2506.20066", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Token merging combining semantic and spatial awareness", "placements": [ { "stage": "Encoding", "category": "Encoding-Oriented Token Compression", "subcategory": "Attention-Aware" } ], "id": "tosa-token-merging-with-spatial-awareness", "stages": [ "Encoding" ], "categories": [ "Encoding-Oriented Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{tosatokenmergingwithspatialawareness2025,\n title = {ToSA: Token Merging with Spatial Awareness},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2506.20066}\n}" }, { "title": "ViCO: A Training Strategy towards Semantic Aware Dynamic High-Resolution", "paper_url": "https://arxiv.org/abs/2510.12793", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Multiple MLP connectors with different compression ratios", "placements": [ { "stage": "Encoding", "category": "Adaptive Resolution", "subcategory": null } ], "id": "vico-a-training-strategy-towards-semantic-aware-dynamic-high-resolution", "stages": [ "Encoding" ], "categories": [ "Adaptive Resolution" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{vicoatrainingstrategytowardssemanticawared2025,\n title = {ViCO: A Training Strategy towards Semantic Aware Dynamic High-Resolution},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2510.12793}\n}" }, { "title": "VidCom2: Video Compression Commander", "paper_url": "https://arxiv.org/abs/2505.14454", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Dynamic compression based on frame uniqueness", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Diversity-Guided" } ], "id": "vidcom2-video-compression-commander", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{vidcom2videocompressioncommander2025,\n title = {VidCom2: Video Compression Commander},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2505.14454}\n}" }, { "title": "Video-XL-2: Towards Very Long-Video Understanding Through Task-Aware KV Sparsification", "paper_url": "https://arxiv.org/abs/2506.19225", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Chunk-based attention with historical context for video processing", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Attention-Guided" } ], "id": "video-xl-2-towards-very-long-video-understanding-through-task-aware-kv-sparsification", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{videoxl2towardsverylongvideounderstandingt2025,\n title = {Video-XL-2: Towards Very Long-Video Understanding Through Task-Aware KV Sparsification},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2506.19225}\n}" }, { "title": "VideoNSA: Native Sparse Attention Scales Video Understanding", "paper_url": "https://arxiv.org/abs/2510.02295", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "End-to-end training with sparse attention preserving dense attention for text", "placements": [ { "stage": "Prefilling", "category": "Sparse Attention", "subcategory": null } ], "id": "videonsa-native-sparse-attention-scales-video-understanding", "stages": [ "Prefilling" ], "categories": [ "Sparse Attention" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{videonsanativesparseattentionscalesvideoun2025,\n title = {VideoNSA: Native Sparse Attention Scales Video Understanding},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2510.02295}\n}" }, { "title": "VideoTree: Adaptive tree-based video representation", "paper_url": "https://openaccess.thecvf.com/content/CVPR2025/html/Wang_VideoTree_Adaptive_Tree-based_Video_Representation_for_LLM_Reasoning_on_Long_CVPR_2025_paper.html", "venue": "CVPR 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Multi-granularity tree-based representation extracting query-relevant details", "placements": [ { "stage": "Encoding", "category": "Keyframe Selection", "subcategory": "Training-Free" } ], "id": "videotree-adaptive-tree-based-video-representation", "stages": [ "Encoding" ], "categories": [ "Keyframe Selection" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{videotreeadaptivetreebasedvideorepresentat2025,\n title = {VideoTree: Adaptive tree-based video representation},\n year = {2025},\n howpublished = {CVPR 2025},\n url = {https://openaccess.thecvf.com/content/CVPR2025/html/Wang_VideoTree_Adaptive_Tree-based_Video_Representation_for_LLM_Reasoning_on_Long_CVPR_2025_paper.html}\n}" }, { "title": "VidKV: Plug-and-Play 1.x-Bit KV Cache Quantization", "paper_url": "https://arxiv.org/abs/2503.16257", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Sub-2-bit quantization with differential treatment for K and V", "placements": [ { "stage": "Decoding", "category": "KV Cache Compression", "subcategory": "Bit-Level" } ], "id": "vidkv-plug-and-play-1-x-bit-kv-cache-quantization", "stages": [ "Decoding" ], "categories": [ "KV Cache Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{vidkvplugandplay1xbitkvcachequantization2025,\n title = {VidKV: Plug-and-Play 1.x-Bit KV Cache Quantization},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2503.16257}\n}" }, { "title": "VisionThink: Smart and efficient vision language model via reinforcement learning", "paper_url": "https://doi.org/10.48550/arXiv.2507.13348", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Dynamically processes distinct samples with different resolutions", "placements": [ { "stage": "Encoding", "category": "Adaptive Resolution", "subcategory": null } ], "id": "visionthink-smart-and-efficient-vision-language-model-via-reinforcement-learning", "stages": [ "Encoding" ], "categories": [ "Adaptive Resolution" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{visionthinksmartandefficientvisionlanguage2025,\n title = {VisionThink: Smart and efficient vision language model via reinforcement learning},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://doi.org/10.48550/arXiv.2507.13348}\n}" }, { "title": "VisionZip: Longer is Better but Not Necessary", "paper_url": "https://arxiv.org/abs/2412.04467", "venue": "CVPR 2025", "year": 2025, "code_url": "https://github.com/dvlab-research/VisionZip", "code_type": "Code", "contribution": "Selects informative tokens using visual attention from encoder", "placements": [ { "stage": "Encoding", "category": "Encoding-Oriented Token Compression", "subcategory": "Attention-Aware" } ], "id": "visionzip-longer-is-better-but-not-necessary", "stages": [ "Encoding" ], "categories": [ "Encoding-Oriented Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{visionziplongerisbetterbutnotnecessary2025,\n title = {VisionZip: Longer is Better but Not Necessary},\n year = {2025},\n howpublished = {CVPR 2025},\n url = {https://arxiv.org/abs/2412.04467}\n}" }, { "title": "ViSpec: Accelerating VLMs with Vision-Aware Speculative Decoding", "paper_url": "https://arxiv.org/abs/2509.15235", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Lightweight vision adaptor to compress image tokens for draft model", "placements": [ { "stage": "Decoding", "category": "Speculative Decoding", "subcategory": "Training-Aware" } ], "id": "vispec-accelerating-vlms-with-vision-aware-speculative-decoding", "stages": [ "Decoding" ], "categories": [ "Speculative Decoding" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{vispecacceleratingvlmswithvisionawarespecu2025,\n title = {ViSpec: Accelerating VLMs with Vision-Aware Speculative Decoding},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2509.15235}\n}" }, { "title": "XAttention: Block sparse attention with antidiagonal scoring", "paper_url": "https://arxiv.org/abs/2503.16428", "venue": "arXiv 2025", "year": 2025, "code_url": null, "code_type": null, "contribution": "Block sparse attention with antidiagonal scoring for efficient block estimation", "placements": [ { "stage": "Prefilling", "category": "Sparse Attention", "subcategory": null } ], "id": "xattention-block-sparse-attention-with-antidiagonal-scoring", "stages": [ "Prefilling" ], "categories": [ "Sparse Attention" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{xattentionblocksparseattentionwithantidiag2025,\n title = {XAttention: Block sparse attention with antidiagonal scoring},\n year = {2025},\n howpublished = {arXiv 2025},\n url = {https://arxiv.org/abs/2503.16428}\n}" }, { "title": "AIM: Adaptive Inference of Multi-Modal LLMs", "paper_url": "https://doi.org/10.48550/ARXIV.2412.03248", "venue": "arXiv 2024", "year": 2024, "code_url": null, "code_type": null, "contribution": "Spatiotemporal token merging to reduce video redundancy", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Diversity-Guided" } ], "id": "aim-adaptive-inference-of-multi-modal-llms", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{aimadaptiveinferenceofmultimodalllms2024,\n title = {AIM: Adaptive Inference of Multi-Modal LLMs},\n year = {2024},\n howpublished = {arXiv 2024},\n url = {https://doi.org/10.48550/ARXIV.2412.03248}\n}" }, { "title": "ConvLLaVA: Hierarchical backbones as visual encoder for large multimodal models", "paper_url": "https://arxiv.org/abs/2405.15738", "venue": "arXiv 2024", "year": 2024, "code_url": "https://github.com/alibaba/conv-llava", "code_type": "Code", "contribution": "Compresses high-resolution images into information-rich visual features", "placements": [ { "stage": "Encoding", "category": "Efficient Vision Encoders", "subcategory": "Image-Related" } ], "id": "convllava-hierarchical-backbones-as-visual-encoder-for-large-multimodal-models", "stages": [ "Encoding" ], "categories": [ "Efficient Vision Encoders" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{convllavahierarchicalbackbonesasvisualenco2024,\n title = {ConvLLaVA: Hierarchical backbones as visual encoder for large multimodal models},\n year = {2024},\n howpublished = {arXiv 2024},\n url = {https://arxiv.org/abs/2405.15738}\n}" }, { "title": "Dynamic-VLM: Simple Dynamic Visual Token Compression", "paper_url": "https://doi.org/10.48550/ARXIV.2412.09530", "venue": "arXiv 2024", "year": 2024, "code_url": null, "code_type": null, "contribution": "Dynamic visual token compression architecture adapting to different lengths", "placements": [ { "stage": "Encoding", "category": "Encoding-Oriented Token Compression", "subcategory": "Attention-Free" }, { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Diversity-Guided" } ], "id": "dynamic-vlm-simple-dynamic-visual-token-compression", "stages": [ "Encoding", "Prefilling" ], "categories": [ "Encoding-Oriented Token Compression", "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{dynamicvlmsimpledynamicvisualtokencompress2024,\n title = {Dynamic-VLM: Simple Dynamic Visual Token Compression},\n year = {2024},\n howpublished = {arXiv 2024},\n url = {https://doi.org/10.48550/ARXIV.2412.09530}\n}" }, { "title": "Dynamic-VLM: Simple Dynamic Visual Token Compression for VideoLLM", "paper_url": "https://doi.org/10.48550/ARXIV.2412.09530", "venue": "arXiv 2024", "year": 2024, "code_url": null, "code_type": null, "contribution": "Dynamic visual token compression architecture adapting to different lengths", "placements": [ { "stage": "Encoding", "category": "Efficient Modality Adapters", "subcategory": null } ], "id": "dynamic-vlm-simple-dynamic-visual-token-compression-for-videollm", "stages": [ "Encoding" ], "categories": [ "Efficient Modality Adapters" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{dynamicvlmsimpledynamicvisualtokencompress2024,\n title = {Dynamic-VLM: Simple Dynamic Visual Token Compression for VideoLLM},\n year = {2024},\n howpublished = {arXiv 2024},\n url = {https://doi.org/10.48550/ARXIV.2412.09530}\n}" }, { "title": "Elastic Cache: Efficient inference of vision instruction-following models", "paper_url": "https://doi.org/10.1007/978-3-031-72643-9_4", "venue": "ECCV 2024", "year": 2024, "code_url": null, "code_type": null, "contribution": "Cache merging strategy fusing less important KVs guided by distinct metrics", "placements": [ { "stage": "Decoding", "category": "KV Cache Compression", "subcategory": "Token-Level" } ], "id": "elastic-cache-efficient-inference-of-vision-instruction-following-models", "stages": [ "Decoding" ], "categories": [ "KV Cache Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{elasticcacheefficientinferenceofvisioninst2024,\n title = {Elastic Cache: Efficient inference of vision instruction-following models},\n year = {2024},\n howpublished = {ECCV 2024},\n url = {https://doi.org/10.1007/978-3-031-72643-9_4}\n}" }, { "title": "FastV: An image is worth 1/2 tokens after layer 2", "paper_url": "https://doi.org/10.1007/978-3-031-73004-7_2", "venue": "ECCV 2024", "year": 2024, "code_url": "https://github.com/pkunlp-icler/FastV", "code_type": "Code", "contribution": "Learns attention patterns in early layers to prune visual tokens in deeper layers", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Attention-Guided" } ], "id": "fastv-an-image-is-worth-1-2-tokens-after-layer-2", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{fastvanimageisworth12tokensafterlayer22024,\n title = {FastV: An image is worth 1/2 tokens after layer 2},\n year = {2024},\n howpublished = {ECCV 2024},\n url = {https://doi.org/10.1007/978-3-031-73004-7_2}\n}" }, { "title": "Frame-Voyager: Learning to query frames for video LLMs", "paper_url": "https://arxiv.org/abs/2410.03226", "venue": "arXiv 2024", "year": 2024, "code_url": null, "code_type": null, "contribution": "Learns to query informative frame combinations", "placements": [ { "stage": "Encoding", "category": "Keyframe Selection", "subcategory": "Training-Aware" } ], "id": "frame-voyager-learning-to-query-frames-for-video-llms", "stages": [ "Encoding" ], "categories": [ "Keyframe Selection" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{framevoyagerlearningtoqueryframesforvideol2024,\n title = {Frame-Voyager: Learning to query frames for video LLMs},\n year = {2024},\n howpublished = {arXiv 2024},\n url = {https://arxiv.org/abs/2410.03226}\n}" }, { "title": "FrameFusion: Combining similarity and importance", "paper_url": "https://arxiv.org/abs/2501.01986", "venue": "arXiv 2024", "year": 2024, "code_url": null, "code_type": null, "contribution": "Merges tokens in shallow layers and prunes in deep layers", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Diversity-Guided" }, { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Attention-Guided" } ], "id": "framefusion-combining-similarity-and-importance", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{framefusioncombiningsimilarityandimportanc2024,\n title = {FrameFusion: Combining similarity and importance},\n year = {2024},\n howpublished = {arXiv 2024},\n url = {https://arxiv.org/abs/2501.01986}\n}" }, { "title": "KeyVideoLLM: Towards large-scale video keyframe selection", "paper_url": "https://arxiv.org/abs/2407.03104", "venue": "arXiv 2024", "year": 2024, "code_url": null, "code_type": null, "contribution": "Employs frozen models as plug-and-play selectors for keyframe localization", "placements": [ { "stage": "Encoding", "category": "Keyframe Selection", "subcategory": "Training-Free" } ], "id": "keyvideollm-towards-large-scale-video-keyframe-selection", "stages": [ "Encoding" ], "categories": [ "Keyframe Selection" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{keyvideollmtowardslargescalevideokeyframes2024,\n title = {KeyVideoLLM: Towards large-scale video keyframe selection},\n year = {2024},\n howpublished = {arXiv 2024},\n url = {https://arxiv.org/abs/2407.03104}\n}" }, { "title": "Less is More: A Simple yet Effective Token Reduction Method", "paper_url": "https://arxiv.org/abs/2409.10994", "venue": "arXiv 2024", "year": 2024, "code_url": null, "code_type": null, "contribution": "Token reduction using both CLIP metric and similarity (TRIM)", "placements": [ { "stage": "Encoding", "category": "Encoding-Oriented Token Compression", "subcategory": "Attention-Free" } ], "id": "less-is-more-a-simple-yet-effective-token-reduction-method", "stages": [ "Encoding" ], "categories": [ "Encoding-Oriented Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{lessismoreasimpleyeteffectivetokenreductio2024,\n title = {Less is More: A Simple yet Effective Token Reduction Method},\n year = {2024},\n howpublished = {arXiv 2024},\n url = {https://arxiv.org/abs/2409.10994}\n}" }, { "title": "Long context transfer from language to vision", "paper_url": "https://arxiv.org/abs/2406.16852", "venue": "arXiv 2024", "year": 2024, "code_url": null, "code_type": null, "contribution": "Vision encoder explicitly trained for long video scenarios", "placements": [ { "stage": "Encoding", "category": "Efficient Vision Encoders", "subcategory": "Video-Related" } ], "id": "long-context-transfer-from-language-to-vision", "stages": [ "Encoding" ], "categories": [ "Efficient Vision Encoders" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{longcontexttransferfromlanguagetovision2024,\n title = {Long context transfer from language to vision},\n year = {2024},\n howpublished = {arXiv 2024},\n url = {https://arxiv.org/abs/2406.16852}\n}" }, { "title": "LongVLM: Efficient long video understanding via large language models", "paper_url": "https://doi.org/10.1007/978-3-031-73414-4_26", "venue": "ECCV 2024", "year": 2024, "code_url": "https://github.com/ziplab/LongVLM", "code_type": "Code", "contribution": "Vision encoder explicitly trained for long video scenarios", "placements": [ { "stage": "Encoding", "category": "Efficient Vision Encoders", "subcategory": "Video-Related" } ], "id": "longvlm-efficient-long-video-understanding-via-large-language-models", "stages": [ "Encoding" ], "categories": [ "Efficient Vision Encoders" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{longvlmefficientlongvideounderstandingvial2024,\n title = {LongVLM: Efficient long video understanding via large language models},\n year = {2024},\n howpublished = {ECCV 2024},\n url = {https://doi.org/10.1007/978-3-031-73414-4_26}\n}" }, { "title": "LOOK-M: Look-once optimization in KV cache", "paper_url": "https://arxiv.org/abs/2406.18139", "venue": "arXiv 2024", "year": 2024, "code_url": null, "code_type": null, "contribution": "Text-prior compression policy prioritizing textual KVs while evicting visual tokens", "placements": [ { "stage": "Decoding", "category": "KV Cache Compression", "subcategory": "Token-Level" } ], "id": "look-m-look-once-optimization-in-kv-cache", "stages": [ "Decoding" ], "categories": [ "KV Cache Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{lookmlookonceoptimizationinkvcache2024,\n title = {LOOK-M: Look-once optimization in KV cache},\n year = {2024},\n howpublished = {arXiv 2024},\n url = {https://arxiv.org/abs/2406.18139}\n}" }, { "title": "MovieChat: From dense token to sparse memory for long video understanding", "paper_url": "https://doi.org/10.1109/CVPR52733.2024.01725", "venue": "CVPR 2024", "year": 2024, "code_url": "https://github.com/wenhaochai/MovieChat", "code_type": "Code", "contribution": "Vision encoder explicitly trained for long video scenarios", "placements": [ { "stage": "Encoding", "category": "Efficient Vision Encoders", "subcategory": "Video-Related" } ], "id": "moviechat-from-dense-token-to-sparse-memory-for-long-video-understanding", "stages": [ "Encoding" ], "categories": [ "Efficient Vision Encoders" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{moviechatfromdensetokentosparsememoryforlo2024,\n title = {MovieChat: From dense token to sparse memory for long video understanding},\n year = {2024},\n howpublished = {CVPR 2024},\n url = {https://doi.org/10.1109/CVPR52733.2024.01725}\n}" }, { "title": "PVC: Progressive Visual Token Compression", "paper_url": "https://arxiv.org/abs/2412.09613", "venue": "arXiv 2024", "year": 2024, "code_url": null, "code_type": null, "contribution": "Progressive compression strategy extending images as static videos", "placements": [ { "stage": "Encoding", "category": "Encoding-Oriented Token Compression", "subcategory": "Attention-Free" } ], "id": "pvc-progressive-visual-token-compression", "stages": [ "Encoding" ], "categories": [ "Encoding-Oriented Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{pvcprogressivevisualtokencompression2024,\n title = {PVC: Progressive Visual Token Compression},\n year = {2024},\n howpublished = {arXiv 2024},\n url = {https://arxiv.org/abs/2412.09613}\n}" }, { "title": "PyramidDrop: Accelerating via pyramid visual redundancy reduction", "paper_url": "https://arxiv.org/abs/2410.17247", "venue": "arXiv 2024", "year": 2024, "code_url": "https://github.com/Cooperx521/PyramidDrop", "code_type": "Code", "contribution": "Multi-stage pruning using attention score ranking", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Attention-Guided" } ], "id": "pyramiddrop-accelerating-via-pyramid-visual-redundancy-reduction", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{pyramiddropacceleratingviapyramidvisualred2024,\n title = {PyramidDrop: Accelerating via pyramid visual redundancy reduction},\n year = {2024},\n howpublished = {arXiv 2024},\n url = {https://arxiv.org/abs/2410.17247}\n}" }, { "title": "Qwen2-VL: Enhancing vision-language model's perception of the world at any resolution", "paper_url": "https://arxiv.org/abs/2409.12191", "venue": "arXiv 2024", "year": 2024, "code_url": "https://github.com/QwenLM/Qwen2-VL", "code_type": "Code", "contribution": "Native Dynamic Resolution framework enabling adaptive visual token generation", "placements": [ { "stage": "Encoding", "category": "Efficient Vision Encoders", "subcategory": "Video-Related" } ], "id": "qwen2-vl-enhancing-vision-language-model-s-perception-of-the-world-at-any-resolution", "stages": [ "Encoding" ], "categories": [ "Efficient Vision Encoders" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{qwen2vlenhancingvisionlanguagemodelspercep2024,\n title = {Qwen2-VL: Enhancing vision-language model's perception of the world at any resolution},\n year = {2024},\n howpublished = {arXiv 2024},\n url = {https://arxiv.org/abs/2409.12191}\n}" }, { "title": "SparseVLM: Visual token sparsification for efficient inference", "paper_url": "https://arxiv.org/abs/2410.04417", "venue": "arXiv 2024", "year": 2024, "code_url": null, "code_type": null, "contribution": "Sparsifies visual tokens based on question prompt through text-visual attention scores", "placements": [ { "stage": "Prefilling", "category": "Token Compression", "subcategory": "Attention-Guided" } ], "id": "sparsevlm-visual-token-sparsification-for-efficient-inference", "stages": [ "Prefilling" ], "categories": [ "Token Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{sparsevlmvisualtokensparsificationforeffic2024,\n title = {SparseVLM: Visual token sparsification for efficient inference},\n year = {2024},\n howpublished = {arXiv 2024},\n url = {https://arxiv.org/abs/2410.04417}\n}" }, { "title": "Video-ChatGPT: Towards detailed video understanding via large vision and language models", "paper_url": "https://doi.org/10.18653/v1/2024.acl-long.679", "venue": "ACL 2024", "year": 2024, "code_url": "https://github.com/mbzuai-oryx/Video-ChatGPT", "code_type": "Code", "contribution": "Applies pooling over visual tokens to obtain compact visual representations", "placements": [ { "stage": "Encoding", "category": "Efficient Vision Encoders", "subcategory": "Video-Related" } ], "id": "video-chatgpt-towards-detailed-video-understanding-via-large-vision-and-language-models", "stages": [ "Encoding" ], "categories": [ "Efficient Vision Encoders" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{videochatgpttowardsdetailedvideounderstand2024,\n title = {Video-ChatGPT: Towards detailed video understanding via large vision and language models},\n year = {2024},\n howpublished = {ACL 2024},\n url = {https://doi.org/10.18653/v1/2024.acl-long.679}\n}" }, { "title": "VILA: Efficient video-language alignment for video question answering", "paper_url": "https://doi.org/10.1007/978-3-031-73033-7_11", "venue": "ECCV 2024", "year": 2024, "code_url": null, "code_type": null, "contribution": "Text-guided Frame-Prompter learning to extract question-related frames", "placements": [ { "stage": "Encoding", "category": "Keyframe Selection", "subcategory": "Training-Aware" } ], "id": "vila-efficient-video-language-alignment-for-video-question-answering", "stages": [ "Encoding" ], "categories": [ "Keyframe Selection" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{vilaefficientvideolanguagealignmentforvide2024,\n title = {VILA: Efficient video-language alignment for video question answering},\n year = {2024},\n howpublished = {ECCV 2024},\n url = {https://doi.org/10.1007/978-3-031-73033-7_11}\n}" }, { "title": "VL-Cache: Sparsity and modality-aware KV cache compression", "paper_url": "https://arxiv.org/abs/2410.23317", "venue": "arXiv 2024", "year": 2024, "code_url": null, "code_type": null, "contribution": "Dynamically sets each layer's cache size according to measured attention sparsity", "placements": [ { "stage": "Decoding", "category": "KV Cache Compression", "subcategory": "Layer-Level" } ], "id": "vl-cache-sparsity-and-modality-aware-kv-cache-compression", "stages": [ "Decoding" ], "categories": [ "KV Cache Compression" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{vlcachesparsityandmodalityawarekvcachecomp2024,\n title = {VL-Cache: Sparsity and modality-aware KV cache compression},\n year = {2024},\n howpublished = {arXiv 2024},\n url = {https://arxiv.org/abs/2410.23317}\n}" }, { "title": "BLIP-2: Bootstrapping Language-Image Pre-training", "paper_url": "https://proceedings.mlr.press/v202/li23q.html", "venue": "ICML 2023", "year": 2023, "code_url": "https://github.com/salesforce/LAVIS", "code_type": "Code", "contribution": "Bridges modality gap with lightweight Querying Transformer (Q-Former)", "placements": [ { "stage": "Encoding", "category": "Efficient Modality Adapters", "subcategory": null } ], "id": "blip-2-bootstrapping-language-image-pre-training", "stages": [ "Encoding" ], "categories": [ "Efficient Modality Adapters" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{blip2bootstrappinglanguageimagepretraining2023,\n title = {BLIP-2: Bootstrapping Language-Image Pre-training},\n year = {2023},\n howpublished = {ICML 2023},\n url = {https://proceedings.mlr.press/v202/li23q.html}\n}" }, { "title": "FastViT: A fast hybrid vision transformer using structural reparameterization", "paper_url": "https://doi.org/10.1109/ICCV51070.2023.00532", "venue": "ICCV 2023", "year": 2023, "code_url": "https://github.com/apple/ml-fastvit", "code_type": "Code", "contribution": "Novel token mixing operators and structural reparameterization", "placements": [ { "stage": "Encoding", "category": "Efficient Vision Encoders", "subcategory": "Image-Related" } ], "id": "fastvit-a-fast-hybrid-vision-transformer-using-structural-reparameterization", "stages": [ "Encoding" ], "categories": [ "Efficient Vision Encoders" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{fastvitafasthybridvisiontransformerusingst2023,\n title = {FastViT: A fast hybrid vision transformer using structural reparameterization},\n year = {2023},\n howpublished = {ICCV 2023},\n url = {https://doi.org/10.1109/ICCV51070.2023.00532}\n}" }, { "title": "SeViLA: Self-chained image-language model for video localization", "paper_url": "http://papers.nips.cc/paper_files/paper/2023/hash/f22a9af8dbb348952b08bd58d4734b50-Abstract-Conference.html", "venue": "NeurIPS 2023", "year": 2023, "code_url": "https://github.com/Yui010206/SeViLA", "code_type": "Code", "contribution": "Uses frozen models as plug-and-play selectors for frame localization", "placements": [ { "stage": "Encoding", "category": "Keyframe Selection", "subcategory": "Training-Free" } ], "id": "sevila-self-chained-image-language-model-for-video-localization", "stages": [ "Encoding" ], "categories": [ "Keyframe Selection" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{sevilaselfchainedimagelanguagemodelforvide2023,\n title = {SeViLA: Self-chained image-language model for video localization},\n year = {2023},\n howpublished = {NeurIPS 2023},\n url = {http://papers.nips.cc/paper_files/paper/2023/hash/f22a9af8dbb348952b08bd58d4734b50-Abstract-Conference.html}\n}" }, { "title": "Video-LLaMA: An instruction-tuned audio-visual language model", "paper_url": "https://arxiv.org/abs/2306.02858", "venue": "arXiv 2023", "year": 2023, "code_url": "https://github.com/DAMO-NLP-SG/Video-LLaMA", "code_type": "Code", "contribution": "Proposes Video Q-Former for multi-modality video comprehension", "placements": [ { "stage": "Encoding", "category": "Efficient Modality Adapters", "subcategory": null } ], "id": "video-llama-an-instruction-tuned-audio-visual-language-model", "stages": [ "Encoding" ], "categories": [ "Efficient Modality Adapters" ], "added_on": "2026-04-08", "added_in": "b688cfc6aea84d46c54338853a8d3c2f00f753a3", "bibtex": "@misc{videollamaaninstructiontunedaudiovisuallan2023,\n title = {Video-LLaMA: An instruction-tuned audio-visual language model},\n year = {2023},\n howpublished = {arXiv 2023},\n url = {https://arxiv.org/abs/2306.02858}\n}" } ] }