schema_version: 1
last_updated: "2026-10-20"
source: "README.md"
repository: "https://github.com/SuDIS-ZJU/Efficient-LVLMs-Inference"
papers:
  - id: "3dzip-spatial-aware-feature-diversity-guided-token-compression-for-3d-question-answering"
    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"
    added_on: "2026-10-20"
    added_in: "162519ad40579f9814e2a0f52443728ea11d6298"
    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: "agilepruner-an-empirical-study-of-attention-and-diversity-for-adaptive-visual-token-pruning-in-large-vision-language-models"
    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"
    added_on: "2026-10-20"
    added_in: "162519ad40579f9814e2a0f52443728ea11d6298"
    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: "avtp-multi-image-visual-token-pruning-in-large-visual-language-models"
    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"
    added_on: "2026-09-21"
    added_in: "fd9fa236feead15f9fc0e8a157c3b78ae56f471d"
    contribution: "Adaptively assigns token budgets across images and supports diverse VLM architectures"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Attention-Guided"
  - id: "coverpruner-who-speaks-for-the-pruned-visual-token-pruning-as-coverage-optimization"
    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
    added_on: "2026-09-21"
    added_in: "fd9fa236feead15f9fc0e8a157c3b78ae56f471d"
    contribution: "Formulates pruning as query-weighted representational coverage maximization"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Diversity-Guided"
  - id: "dstp-why-and-when-visual-token-pruning-fails"
    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
    added_on: "2026-09-21"
    added_in: "fd9fa236feead15f9fc0e8a157c3b78ae56f471d"
    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: "evocomp-learning-visual-token-compression-via-semantic-guided-evolutionary-labeling"
    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
    added_on: "2026-09-21"
    added_in: "fd9fa236feead15f9fc0e8a157c3b78ae56f471d"
    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: "hi-lo-prune-look-at-what-you-ll-lose-before-pruning-with-hierarchical-token-selection"
    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"
    added_on: "2026-10-20"
    added_in: "bf8daaed7993c2258fb8dc983ed32c4964e94ce9"
    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: "hybridkv-hybrid-kv-cache-compression-for-efficient-multimodal-large-language-model-inference"
    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
    added_on: "2026-09-21"
    added_in: "fd9fa236feead15f9fc0e8a157c3b78ae56f471d"
    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: "icctp-instruction-guided-cross-modal-clustering-for-training-free-visual-token-pruning"
    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
    added_on: "2026-09-21"
    added_in: "fd9fa236feead15f9fc0e8a157c3b78ae56f471d"
    contribution: "Uses instruction noun anchors for cross-modal clustering and high-ratio token pruning"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Attention-Guided"
  - id: "jag-direct-box-prediction-for-efficient-visual-grounding"
    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"
    added_on: "2026-10-20"
    added_in: "b91d7270a30bcc4659bbbfb7e9c30688af5b570a"
    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: "mmspec-benchmarking-speculative-decoding-for-vision-language-models"
    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"
    added_on: "2026-09-21"
    added_in: "fd9fa236feead15f9fc0e8a157c3b78ae56f471d"
    contribution: "Benchmarks ten speculative decoding methods and introduces vision-adaptive ViSkip"
    placements:
      - stage: "Decoding"
        category: "Speculative Decoding"
        subcategory: "Training-Free"
  - id: "occamtoken-efficient-vlm-inference-with-training-free-and-budget-adaptive-token-pruning"
    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
    added_on: "2026-09-21"
    added_in: "fd9fa236feead15f9fc0e8a157c3b78ae56f471d"
    contribution: "Uses register-anchored relative evidence tests for image- and query-adaptive pruning"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Attention-Guided"
  - id: "q-cache-visual-attention-is-valuable-in-less-than-half-of-decode-layers-for-multimodal-large-language-model"
    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
    added_on: "2026-09-21"
    added_in: "fd9fa236feead15f9fc0e8a157c3b78ae56f471d"
    contribution: "Reuses similar attention queries across adjacent layers through a lightweight layer-shared cache"
    placements:
      - stage: "Decoding"
        category: "Attention Reuse"
        subcategory: null
  - id: "semantic-guided-slow-fast-pruning-of-visual-tokens-for-vision-language-models"
    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
    added_on: "2026-09-21"
    added_in: "83bb2d56348b3c0ba8a8c18c435389561cac1c10"
    contribution: "Semantic-guided slow-fast pruning of visual tokens for VLMs"
    placements:
      - stage: "Encoding"
        category: "Encoding-Oriented Token Compression"
        subcategory: "Attention-Aware"
  - id: "sieve-when-vision-becomes-text"
    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
    added_on: "2026-09-21"
    added_in: "fd9fa236feead15f9fc0e8a157c3b78ae56f471d"
    contribution: "Retains visual information unexplained by the text subspace using cross-modal residual guidance"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Attention-Guided"
  - id: "stc-accelerating-streaming-video-large-language-models-via-hierarchical-token-compression"
    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"
    added_on: "2026-09-21"
    added_in: "fd9fa236feead15f9fc0e8a157c3b78ae56f471d"
    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: "swiftvlm-efficient-vision-language-model-inference-via-cross-layer-token-bypass"
    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
    added_on: "2026-09-21"
    added_in: "fd9fa236feead15f9fc0e8a157c3b78ae56f471d"
    contribution: "Bypasses rather than permanently discards unselected tokens, enabling later-layer re-evaluation"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Attention-Guided"
  - id: "unified-spatiotemporal-token-compression-for-video-llms-at-ultra-low-retention"
    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
    added_on: "2026-09-21"
    added_in: "fd9fa236feead15f9fc0e8a157c3b78ae56f471d"
    contribution: "Globally allocates a unified spatiotemporal token budget using contribution and redundancy signals"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Diversity-Guided"
  - id: "adaptive-fast-and-slow-visual-program-reasoning-for-long-form-videoqa"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    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: "aircache-activating-inter-modal-relevancy-kv-cache-compression"
    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
    added_on: "2026-09-21"
    added_in: "fd9fa236feead15f9fc0e8a157c3b78ae56f471d"
    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: "akvq-vl-attention-aware-kv-cache-adaptive-2-bit-quantization"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    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: "atp-llava-adaptive-token-pruning-for-lvlms"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Learnable adaptive token pruning module computing importance score"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Attention-Guided"
  - id: "beyond-text-visual-attention-exploiting-visual-cues-for-effective-token-pruning"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Selects informative tokens using visual attention (VisPruner)"
    placements:
      - stage: "Encoding"
        category: "Encoding-Oriented Token Compression"
        subcategory: "Attention-Free"
  - id: "btp-balanced-token-pruning"
    title: "BTP: Balanced Token Pruning"
    paper_url: "https://arxiv.org/abs/2505.22038"
    venue: "arXiv 2025"
    year: 2025
    code_url: null
    code_type: null
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Multi-stage pruning with diversity and attention ranking objectives"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Attention-Guided"
  - id: "calibquant-1-bit-kv-cache-quantization-for-multimodal-llms"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Channel-wise 1-bit quantization with post-calibration for extreme values"
    placements:
      - stage: "Decoding"
        category: "KV Cache Compression"
        subcategory: "Bit-Level"
  - id: "cdpruner-maximizing-conditional-diversity-for-token-pruning"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Determinantal Point Processes (DPP) maximizing conditional diversity"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Diversity-Guided"
  - id: "dart-stop-looking-for-important-tokens-duplication-matters-more"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Pivot-based duplication pruning selecting tokens with low duplication"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Diversity-Guided"
  - id: "divprune-diversity-based-visual-token-pruning"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Max-Min diversity optimization for token subset selection"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Diversity-Guided"
  - id: "dycoke-dynamic-compression-of-tokens-for-fast-video-llms"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Plug-and-play temporal compression module minimizing temporal redundancy"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Diversity-Guided"
  - id: "fastcache-optimizing-multimodal-llm-serving"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Lightweight modality-specific compressor learning compression patterns"
    placements:
      - stage: "Decoding"
        category: "KV Cache Compression"
        subcategory: "Token-Level"
  - id: "fastvid-dynamic-density-pruning-for-fast-video-llms"
    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"
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Temporal segmentation and density spatiotemporal pruning"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Diversity-Guided"
  - id: "fastvlm-efficient-vision-encoding-for-vision-language-models"
    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"
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Hybrid vision encoder outputting fewer tokens and reducing encoding time"
    placements:
      - stage: "Encoding"
        category: "Efficient Vision Encoders"
        subcategory: "Image-Related"
  - id: "fastvlm-self-speculative-decoding-for-fast-vision-language-model-inference"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Imitation-based draft model learning from deeper representations (self-speculative decoding)"
    placements:
      - stage: "Decoding"
        category: "Speculative Decoding"
        subcategory: "Training-Aware"
  - id: "fit-and-prune-fast-and-training-free-visual-token-pruning"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Minimizes divergence of attention distributions before and after pruning"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Attention-Guided"
  - id: "flash-latent-aware-semi-autoregressive-speculative-decoding"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Visual token compression mechanism and semi-autoregressive head for draft model optimization"
    placements:
      - stage: "Decoding"
        category: "Speculative Decoding"
        subcategory: "Training-Aware"
  - id: "focus-efficient-keyframe-selection-for-long-video-understanding"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Formulates keyframe selection as combinatorial pure-exploration"
    placements:
      - stage: "Encoding"
        category: "Keyframe Selection"
        subcategory: "Training-Free"
  - id: "folder-accelerating-multi-modal-large-language-models"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    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: "g-prune-training-free-visual-token-pruning-from-graph-perspective"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Similarity graph and information flow to retain representative tokens"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Diversity-Guided"
  - id: "glyph-scaling-context-windows-via-visual-text-compression"
    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"
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    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: "hivtp-hierarchical-visual-token-pruning"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Attention maps from middle encoder layers to estimate visual token importance"
    placements:
      - stage: "Encoding"
        category: "Encoding-Oriented Token Compression"
        subcategory: "Attention-Aware"
  - id: "holitom-holistic-token-merging-for-fast-video-llms"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Global redundancy-aware segmentation followed by spatiotemporal merging"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Diversity-Guided"
  - id: "infinipot-v-memory-constrained-kv-cache-compression"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    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: "livevlm-efficient-online-video-understanding-via-streaming-oriented-kv-cache"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    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: "llava-prumerge-adaptive-token-reduction"
    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"
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Reduces visual tokens according to similarities between class and spatial tokens"
    placements:
      - stage: "Encoding"
        category: "Encoding-Oriented Token Compression"
        subcategory: "Attention-Free"
  - id: "longvu-spatiotemporal-adaptive-compression"
    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"
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Preserves full features for query-relevant frames"
    placements:
      - stage: "Encoding"
        category: "Adaptive Resolution"
        subcategory: null
  - id: "longvu-spatiotemporal-adaptive-compression-for-long-video-language-understanding"
    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"
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Preserves full features for query-relevant frames while applying spatial pooling"
    placements:
      - stage: "Encoding"
        category: "Efficient Vision Encoders"
        subcategory: "Video-Related"
  - id: "m-llm-based-video-frame-selection-for-efficient-video-understanding"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Uses spatial and temporal signals as supervision to train frame selector"
    placements:
      - stage: "Encoding"
        category: "Keyframe Selection"
        subcategory: "Training-Aware"
  - id: "madakv-adaptive-modality-perception-kv-cache-eviction"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Inter-layer compensation mechanism dynamically adjusting budgets"
    placements:
      - stage: "Decoding"
        category: "KV Cache Compression"
        subcategory: "Layer-Level"
  - id: "massv-multimodal-adaptation-and-self-data-distillation-for-speculative-decoding-of-vlms"
    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
    added_on: "2026-09-21"
    added_in: "fd9fa236feead15f9fc0e8a157c3b78ae56f471d"
    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: "meda-dynamic-kv-cache-allocation-for-efficient-multimodal-inference"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Cross-modal attention entropy guiding cache allocation to layers with complex interactions"
    placements:
      - stage: "Decoding"
        category: "KV Cache Compression"
        subcategory: "Layer-Level"
  - id: "mminference-modality-aware-permutation-sparse-attention"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Identifies three distinct attention patterns in LVLMs with modality-aware permutation"
    placements:
      - stage: "Prefilling"
        category: "Sparse Attention"
        subcategory: null
  - id: "msd-speculative-decoding-reimagined-for-multimodal-large-language-models"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Two-stage training enabling draft model to acquire language modeling and visual perception"
    placements:
      - stage: "Decoding"
        category: "Speculative Decoding"
        subcategory: "Training-Aware"
  - id: "pixelthink-towards-efficient-chain-of-pixel-reasoning"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Reinforcement learning to regulate reasoning chain length based on task difficulty and model confidence"
    placements:
      - stage: "Decoding"
        category: "Efficient Reasoning"
        subcategory: null
  - id: "prolonged-reasoning-is-not-all-you-need-certainty-based-adaptive-routing"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Certainty-based routing triggering long thought chains only when initial answer exhibits high uncertainty"
    placements:
      - stage: "Decoding"
        category: "Efficient Reasoning"
        subcategory: null
  - id: "prunevid-visual-token-pruning-for-efficient-video-llms"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Spatiotemporal token merging before LLMs"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Diversity-Guided"
  - id: "q-frame-query-aware-frame-selection-and-multi-resolution-adaptation"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    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: "rekv-streaming-video-qa-with-in-context-video-kv-cache-retrieval"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Retrieval-based framework offloading video chunks to external memory"
    placements:
      - stage: "Decoding"
        category: "KV Cache Compression"
        subcategory: "Token-Level"
  - id: "spargeattn-accurate-sparse-attention-accelerating-any-model-inference"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Two-stage online filter to skip unimportant regions in sparse attention"
    placements:
      - stage: "Prefilling"
        category: "Sparse Attention"
        subcategory: null
  - id: "sparsemm-head-sparsity-emerges-from-visual-concept-responses"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Identifies vital visual heads and allocates asymmetric budgets based on visual relevance"
    placements:
      - stage: "Decoding"
        category: "KV Cache Compression"
        subcategory: "Head-Level"
  - id: "sparsevila-decoupling-visual-sparsity-for-efficient-vlm-inference"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Estimates token importance from visual encoder's self-attention maps"
    placements:
      - stage: "Encoding"
        category: "Encoding-Oriented Token Compression"
        subcategory: "Attention-Aware"
  - id: "spec-llava-accelerating-vlms-with-dynamic-tree-based-speculative-decoding"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Compact distilled draft model paired with tree-based verification algorithm"
    placements:
      - stage: "Decoding"
        category: "Speculative Decoding"
        subcategory: "Training-Aware"
  - id: "specvlm-enhancing-speculative-decoding-via-verifier-guided-token-pruning"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    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-fast-speculative-decoding-in-vision-language-models"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Elastic visual compressor adaptively selecting from multiple compression primitives"
    placements:
      - stage: "Decoding"
        category: "Speculative Decoding"
        subcategory: "Training-Aware"
  - id: "st3-accelerating-mllm-by-spatial-temporal-visual-token-trimming"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    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: "streamingtom-streaming-token-compression"
    title: "StreamingTOM: Streaming Token Compression"
    paper_url: "https://arxiv.org/abs/2510.18269"
    venue: "arXiv 2025"
    year: 2025
    code_url: null
    code_type: null
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Causal temporal reduction with fixed per-frame budget"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Diversity-Guided"
  - id: "streamingvlm-real-time-understanding-for-infinite-video-streams"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Maintains compact token subset by reusing attention sinks and recent token windows"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Attention-Guided"
  - id: "sttm-multi-granular-spatio-temporal-token-merging"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Quadtree spatial transformation with directed pairwise merging"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Diversity-Guided"
  - id: "timechat-online-80-visual-tokens-are-naturally-redundant"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Differential token drop module filtering redundant content in streaming videos"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Diversity-Guided"
  - id: "tokenpacker-efficient-visual-projector-for-multimodal-llm"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Coarse-to-fine scheme injecting enriched characteristics"
    placements:
      - stage: "Encoding"
        category: "Efficient Modality Adapters"
        subcategory: null
  - id: "tosa-token-merging-with-spatial-awareness"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Token merging combining semantic and spatial awareness"
    placements:
      - stage: "Encoding"
        category: "Encoding-Oriented Token Compression"
        subcategory: "Attention-Aware"
  - id: "vico-a-training-strategy-towards-semantic-aware-dynamic-high-resolution"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Multiple MLP connectors with different compression ratios"
    placements:
      - stage: "Encoding"
        category: "Adaptive Resolution"
        subcategory: null
  - id: "vidcom2-video-compression-commander"
    title: "VidCom2: Video Compression Commander"
    paper_url: "https://arxiv.org/abs/2505.14454"
    venue: "arXiv 2025"
    year: 2025
    code_url: null
    code_type: null
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Dynamic compression based on frame uniqueness"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Diversity-Guided"
  - id: "video-xl-2-towards-very-long-video-understanding-through-task-aware-kv-sparsification"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Chunk-based attention with historical context for video processing"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Attention-Guided"
  - id: "videonsa-native-sparse-attention-scales-video-understanding"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "End-to-end training with sparse attention preserving dense attention for text"
    placements:
      - stage: "Prefilling"
        category: "Sparse Attention"
        subcategory: null
  - id: "videotree-adaptive-tree-based-video-representation"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Multi-granularity tree-based representation extracting query-relevant details"
    placements:
      - stage: "Encoding"
        category: "Keyframe Selection"
        subcategory: "Training-Free"
  - id: "vidkv-plug-and-play-1-x-bit-kv-cache-quantization"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Sub-2-bit quantization with differential treatment for K and V"
    placements:
      - stage: "Decoding"
        category: "KV Cache Compression"
        subcategory: "Bit-Level"
  - id: "visionthink-smart-and-efficient-vision-language-model-via-reinforcement-learning"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Dynamically processes distinct samples with different resolutions"
    placements:
      - stage: "Encoding"
        category: "Adaptive Resolution"
        subcategory: null
  - id: "visionzip-longer-is-better-but-not-necessary"
    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"
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Selects informative tokens using visual attention from encoder"
    placements:
      - stage: "Encoding"
        category: "Encoding-Oriented Token Compression"
        subcategory: "Attention-Aware"
  - id: "vispec-accelerating-vlms-with-vision-aware-speculative-decoding"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Lightweight vision adaptor to compress image tokens for draft model"
    placements:
      - stage: "Decoding"
        category: "Speculative Decoding"
        subcategory: "Training-Aware"
  - id: "xattention-block-sparse-attention-with-antidiagonal-scoring"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Block sparse attention with antidiagonal scoring for efficient block estimation"
    placements:
      - stage: "Prefilling"
        category: "Sparse Attention"
        subcategory: null
  - id: "aim-adaptive-inference-of-multi-modal-llms"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Spatiotemporal token merging to reduce video redundancy"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Diversity-Guided"
  - id: "convllava-hierarchical-backbones-as-visual-encoder-for-large-multimodal-models"
    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"
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Compresses high-resolution images into information-rich visual features"
    placements:
      - stage: "Encoding"
        category: "Efficient Vision Encoders"
        subcategory: "Image-Related"
  - id: "dynamic-vlm-simple-dynamic-visual-token-compression"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    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-for-videollm"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Dynamic visual token compression architecture adapting to different lengths"
    placements:
      - stage: "Encoding"
        category: "Efficient Modality Adapters"
        subcategory: null
  - id: "elastic-cache-efficient-inference-of-vision-instruction-following-models"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Cache merging strategy fusing less important KVs guided by distinct metrics"
    placements:
      - stage: "Decoding"
        category: "KV Cache Compression"
        subcategory: "Token-Level"
  - id: "fastv-an-image-is-worth-1-2-tokens-after-layer-2"
    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"
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Learns attention patterns in early layers to prune visual tokens in deeper layers"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Attention-Guided"
  - id: "frame-voyager-learning-to-query-frames-for-video-llms"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Learns to query informative frame combinations"
    placements:
      - stage: "Encoding"
        category: "Keyframe Selection"
        subcategory: "Training-Aware"
  - id: "framefusion-combining-similarity-and-importance"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    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: "keyvideollm-towards-large-scale-video-keyframe-selection"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Employs frozen models as plug-and-play selectors for keyframe localization"
    placements:
      - stage: "Encoding"
        category: "Keyframe Selection"
        subcategory: "Training-Free"
  - id: "less-is-more-a-simple-yet-effective-token-reduction-method"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Token reduction using both CLIP metric and similarity (TRIM)"
    placements:
      - stage: "Encoding"
        category: "Encoding-Oriented Token Compression"
        subcategory: "Attention-Free"
  - id: "long-context-transfer-from-language-to-vision"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    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"
    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"
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Vision encoder explicitly trained for long video scenarios"
    placements:
      - stage: "Encoding"
        category: "Efficient Vision Encoders"
        subcategory: "Video-Related"
  - id: "look-m-look-once-optimization-in-kv-cache"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Text-prior compression policy prioritizing textual KVs while evicting visual tokens"
    placements:
      - stage: "Decoding"
        category: "KV Cache Compression"
        subcategory: "Token-Level"
  - id: "moviechat-from-dense-token-to-sparse-memory-for-long-video-understanding"
    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"
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Vision encoder explicitly trained for long video scenarios"
    placements:
      - stage: "Encoding"
        category: "Efficient Vision Encoders"
        subcategory: "Video-Related"
  - id: "pvc-progressive-visual-token-compression"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Progressive compression strategy extending images as static videos"
    placements:
      - stage: "Encoding"
        category: "Encoding-Oriented Token Compression"
        subcategory: "Attention-Free"
  - id: "pyramiddrop-accelerating-via-pyramid-visual-redundancy-reduction"
    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"
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Multi-stage pruning using attention score ranking"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Attention-Guided"
  - id: "qwen2-vl-enhancing-vision-language-model-s-perception-of-the-world-at-any-resolution"
    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"
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Native Dynamic Resolution framework enabling adaptive visual token generation"
    placements:
      - stage: "Encoding"
        category: "Efficient Vision Encoders"
        subcategory: "Video-Related"
  - id: "sparsevlm-visual-token-sparsification-for-efficient-inference"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Sparsifies visual tokens based on question prompt through text-visual attention scores"
    placements:
      - stage: "Prefilling"
        category: "Token Compression"
        subcategory: "Attention-Guided"
  - id: "video-chatgpt-towards-detailed-video-understanding-via-large-vision-and-language-models"
    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"
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Applies pooling over visual tokens to obtain compact visual representations"
    placements:
      - stage: "Encoding"
        category: "Efficient Vision Encoders"
        subcategory: "Video-Related"
  - id: "vila-efficient-video-language-alignment-for-video-question-answering"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Text-guided Frame-Prompter learning to extract question-related frames"
    placements:
      - stage: "Encoding"
        category: "Keyframe Selection"
        subcategory: "Training-Aware"
  - id: "vl-cache-sparsity-and-modality-aware-kv-cache-compression"
    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
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Dynamically sets each layer's cache size according to measured attention sparsity"
    placements:
      - stage: "Decoding"
        category: "KV Cache Compression"
        subcategory: "Layer-Level"
  - id: "blip-2-bootstrapping-language-image-pre-training"
    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"
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Bridges modality gap with lightweight Querying Transformer (Q-Former)"
    placements:
      - stage: "Encoding"
        category: "Efficient Modality Adapters"
        subcategory: null
  - id: "fastvit-a-fast-hybrid-vision-transformer-using-structural-reparameterization"
    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"
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Novel token mixing operators and structural reparameterization"
    placements:
      - stage: "Encoding"
        category: "Efficient Vision Encoders"
        subcategory: "Image-Related"
  - id: "sevila-self-chained-image-language-model-for-video-localization"
    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"
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Uses frozen models as plug-and-play selectors for frame localization"
    placements:
      - stage: "Encoding"
        category: "Keyframe Selection"
        subcategory: "Training-Free"
  - id: "video-llama-an-instruction-tuned-audio-visual-language-model"
    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"
    added_on: "2026-04-08"
    added_in: "b688cfc6aea84d46c54338853a8d3c2f00f753a3"
    contribution: "Proposes Video Q-Former for multi-modality video comprehension"
    placements:
      - stage: "Encoding"
        category: "Efficient Modality Adapters"
        subcategory: null