mlsanityData to 7 Oct 2026

Visual token pruning and efficiency

635 papers5 this weekHigh confidence in this grouping: the papers form one clear topicHigh confidence in this grouping: the papers form one clear topic · About

Keywords: tokens, token, pruning, visual, compression

Key facts. Visual token pruning and efficiency: 635 papers on arXiv since 2022 (data through 7 Oct 2026). 158 of its 635 mapped papers (25%) cite “An Image is Worth 1/2 Tokens After Layer 2: Plug-and-Play Inference Acceleration for Large Vision-Language Models” (2024, arXiv 2403.06764). 123 of its 635 mapped papers (19%) cite “LLaVA-PruMerge: Adaptive Token Reduction for Efficient Large Multimodal Models” (2024, arXiv 2403.15388). 43 new papers 17 Sep–7 Oct.

What’s new

17 Sep – 7 Oct 2026

43 new papers, 17 Sep – 7 Oct; 5 this week.

  1. This week · 5
  2. Why VLMs Miss Small Objects, and When Zooming In Is Safe also fits: VLM benchmarks and visual reasoning
  3. Adaptive Visual Token Reduction for Accelerated Image Understanding
  4. DIPrune: Task-Aware Token Pruning with Dual Importance for Efficient Multimodal Language Models
  5. VisionWeave: Weaving Elastic Visual Representations as a Native Capability of MLLMs also fits: VLM benchmarks and visual reasoning
  6. Foveated Compression: Selective High-Resolution Preservation for Token-Efficient VLMs
  7. Earlier · 38
  8. When and What to Prune? Stage-Aware Visual Token Pruning for Efficient VLA
  9. FlashGaze: Training-Free Multi-Scale Patch Pruning For Efficient Video Understanding also fits: Video understanding with MLLMs
  10. From Patching to Pruning Visual Computation in Vision Language Models
  11. VETO: Video Efficient Token Optimization for Vision Language Models
  12. Architectural Sampling: Test-Time Scaling via Computational Diversity in Frozen Vision-Language Models also fits: Prompt learning and adaptation of CLIP
All 43
  1. MWOP: Modality-aware Width-wise Operation Pruning for Efficient MLLMs
  2. Fewer Tokens, More Self-Teaching: On-Policy Self-Distillation for Extreme Visual Token Reduction
  3. HAWK: Rethinking Multimodal Drafting for Speculative Decoding
  4. CoVisco: Codec-Native Vision Encoder with Native Token Compression for Unified Image-Video Understanding
  5. Feature-Aware Token Attack for Compression-Triggered Stealthy Failures in Large Vision-Language Models also fits: MLLM jailbreaks and adversarial robustness
  6. Representation Dynamics Reveal Semantic Saliency and Similarity for Visual Token Pruning in MLLMs
  7. MiCo: Mutual Information Coverage Optimization through Semantic Erasure Modeling for Efficient MLLM Inference also fits: MLLM training and visual instruction tuning
  8. TReVS: Integrating Textual Relevance and Visual Saliency for Efficient Vision-Language Model Token Pruning
  9. Hierarchical Compression of Vision-Language Model Benchmarks
  10. Why MLLMs Struggle to Count: Overcoming Individuation and Aggregation Bottlenecks with ConvStack also fits: MLLM training and visual instruction tuning
  11. Task-Oriented Visual Feature Compression via Residual Vector Quantization for Device-Edge Multimodal Inference
  12. Beyond Selection: Token Parameterization for Extreme Visual Token Compression
  13. SPIDER: Multi-Layer Semantic Token Pruning and Adaptive Sub-Layer Skipping in Multimodal Large Language Models
  14. Just MLPs: Efficient Visual State Reconstruction for Multimodal Language Models
  15. Resolution as a First-Class Decision: Task-Conditioned Routing for Efficient Multimodal Large Language Models
  16. SubRot: Signed Gradient Subspace Calibration for VLM Rotation Quantization also fits: Prompt learning and adaptation of CLIP
  17. P4Q: Co-designing Token Pruning and Quantization for Vision-Language Model Acceleration
  18. When Text Matters: Design Principles for Visual Token Pruning in Vision-Language Model
  19. Beyond Reconstruction Loss in Post-Training Quantization: Balanced Fitting for Large Vision-Language Models
  20. ACPruner: Visual Token Pruning as Biased Attention Coverage Maximization in LVLMs
  21. Mind the Spike: Mechanisms and Brittleness of Visual Massive Activations in Large Vision-Language Models also fits: Hallucination mitigation in LVLMs
  22. Readout is not Recovery: Dissociating Coordinate Emission from Visual-Corruption Repair in Vision-Language Models
  23. EAServe: Encode-Aware Disaggregated Serving for Multimodal Large Language Models also fits: MLLM training and visual instruction tuning
  24. From Token Importance to Conditional Removability: Rethinking Visual Token Pruning in Multimodal Large Language Models
  25. Shallow to Deep: Aligning Token Pruning with Stage-wise Roles in LVLMs
  26. RGSQ: Riemannian Geometry-Sensitive Quantization for Large Vision-Language Models
  27. SPHQuant: Efficient extreme low bit weight quantization for Vision-Language Models
  28. VPRune: Efficient Training-free Pre-LLM Visual Token Pruning
  29. Can't Find Waldo: Evaluating VLMs' Sensitivity to Image Resolution and Detail Level also fits: VLM benchmarks and visual reasoning
  30. Layer-Aware Position Embeddings for Visual Token Pruning in Multimodal Large Language Models
  31. HeadGuard: Selective Head Protection for Low-Bit VLM KV-Cache Quantization
  32. QCPruner: Query-Conditioned Population Coverage for Visual Token Pruning
  33. MiX: Micro-Inverted-Scaling for End-to-End Low-Bit Vision-Language Model Acceleration

Foundations

Most cited by papers here

  1. An Image is Worth 1/2 Tokens After Layer 2: Plug-and-Play Inference Acceleration for Large Vision-Language Models2024cited by 158 here
  2. LLaVA-PruMerge: Adaptive Token Reduction for Efficient Large Multimodal Models2024cited by 123 here
  3. SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference2024cited by 119 here
  4. PyramidDrop: Accelerating Your Large Vision-Language Models via Pyramid Visual Redundancy Reduction2024cited by 93 here
  5. VisionZip: Longer is Better but Not Necessary in Vision Language Models2024cited by 88 here

Lineage

The papers this direction builds on, oldest first.

Each paper on the line cites the one before.

  1. 2022Flamingo: a Visual Language Model for Few-Shot Learningcited by 91 here
  2. 2023InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuningcited by 90 here
  3. 2023Evaluating Object Hallucination in Large Vision-Language Modelscited by 145 here
  4. 2023MME: A Comprehensive Evaluation Benchmark for Multimodal Large Language Modelscited by 144 here
  5. 2023SEED-Bench: Benchmarking Multimodal LLMs with Generative Comprehensioncited by 54 here
  6. 2023MM-Vet: Evaluating Large Multimodal Models for Integrated Capabilitiescited by 79 here
  7. 2024An Image is Worth 1/2 Tokens After Layer 2: Plug-and-Play Inference Acceleration for Large Vision-Language Modelscited by 158 here
  8. 2024SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inferencecited by 119 here
  9. 2024PyramidDrop: Accelerating Your Large Vision-Language Models via Pyramid Visual Redundancy Reductioncited by 93 here
  10. 2024VisionZip: Longer is Better but Not Necessary in Vision Language Modelscited by 88 here
14 more it builds on, not in this chain
  1. 2023MMBench: Is Your Multi-modal Model an All-around Player?cited by 169 here
  2. 2023Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyondcited by 80 here
  3. 2023Improved Baselines with Visual Instruction Tuningcited by 175 here
  4. 2023Video-LLaVA: Learning United Visual Representation by Alignment Before Projectioncited by 68 here
  5. 2023InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Taskscited by 92 here
  6. 2024LLaVA-PruMerge: Adaptive Token Reduction for Efficient Large Multimodal Modelscited by 123 here
  7. 2024How Far Are We to GPT-4V? Closing the Gap to Commercial Multimodal Models with Open-Source Suitescited by 47 here
  8. 2024Boosting Multimodal Large Language Models with Visual Tokens Withdrawal for Rapid Inferencecited by 63 here
  9. 2024Video-MME: The First-Ever Comprehensive Evaluation Benchmark of Multi-modal LLMs in Video Analysiscited by 49 here
  10. 2024TokenPacker: Efficient Visual Projector for Multimodal LLMcited by 49 here
  11. 2024LLaVA-OneVision: Easy Visual Task Transfercited by 93 here
  12. 2024Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolutioncited by 114 here
  13. 2024Beyond Text-Visual Attention: Exploiting Visual Cues for Effective Token Pruning in VLMscited by 57 here
  14. 2025DivPrune: Diversity-based Visual Token Pruning for Large Multimodal Modelscited by 59 here

The 24 most cited of 1,041.

Browse

30 of 635, ranked 6–35

  1. 6Boosting Multimodal Large Language Models with Visual Tokens Withdrawal for Rapid Inference202463
  2. 7DivPrune: Diversity-based Visual Token Pruning for Large Multimodal Models202559
  3. 8Beyond Text-Visual Attention: Exploiting Visual Cues for Effective Token Pruning in VLMs202457
  4. 9TokenPacker: Efficient Visual Projector for Multimodal LLM202449
  5. 10Stop Looking for Important Tokens in Multimodal Language Models: Duplication Matters More202545
  6. 11HiRED: Attention-Guided Token Dropping for Efficient Inference of High-Resolution Vision-Language Models202441
  7. 12Fit and Prune: Fast and Training-free Visual Token Pruning for Multi-modal Large Language Models202440
  8. 13ATP-LLaVA: Adaptive Token Pruning for Large Vision Language Models202432
  9. 14Beyond Attention or Similarity: Maximizing Conditional Diversity for Token Pruning in MLLMs202530
  10. 15Dynamic-LLaVA: Efficient Multimodal Large Language Models via Dynamic Vision-language Context Sparsification202427
All 30
  1. 16CrossGET: Cross-Guided Ensemble of Tokens for Accelerating Vision-Language Transformers202326
  2. 17LLaVA-Mini: Efficient Image and Video Large Multimodal Models with One Vision Token202525
  3. 18VoCo-LLaMA: Towards Vision Compression with Large Language Models202425
  4. 19LLaVA-UHD: an LMM Perceiving Any Aspect Ratio and High-Resolution Images202423
  5. 20Multi-Stage Vision Token Dropping: Towards Efficient Multimodal Large Language Model202423
  6. 21TopV: Compatible Token Pruning with Inference Time Optimization for Fast and Low-Memory Multimodal Vision Language Model202522
  7. 22LOOK-M: Look-Once Optimization in KV Cache for Efficient Multimodal Long-Context Inference202421
  8. 23MobileVLM : A Fast, Strong and Open Vision Language Assistant for Mobile Devices202321
  9. 24Token Pruning in Multimodal Large Language Models: Are We Solving the Right Problem?202520
  10. 25Don't Just Chase "Highlighted Tokens" in MLLMs: Revisiting Visual Holistic Context Retention202519
  11. 26Matryoshka Multimodal Models202419
  12. 27Efficient Large Multi-modal Models via Visual Context Compression202419
  13. 28Q-VLM: Post-training Quantization for Large Vision-Language Models202419
  14. 29FastVLM: Efficient Vision Encoding for Vision Language Models202419
  15. 30MBQ: Modality-Balanced Quantization for Large Vision-Language Models202419
  16. 31DeCo: Decoupling Token Compression from Semantic Abstraction in Multimodal Large Language Models202418
  17. 32PACT: Pruning and Clustering-Based Token Reduction for Faster Visual Language Models202517
  18. 33SmolVLM: Redefining small and efficient multimodal models202514
  19. 34ZipVL: Efficient Large Vision-Language Models with Dynamic Token Sparsification202414
  20. 35VL-Cache: Sparsity and Modality-Aware KV Cache Compression for Vision-Language Model Inference Acceleration202414