mlsanityData to 7 Oct 2026

Token pruning and merging for ViTs

108 papers2 this weekHigh confidence in this grouping: the papers form one clear topicHigh confidence in this grouping: the papers form one clear topic · About

Keywords: token, vision, pruning, transformers, merging

Key facts. Token pruning and merging for ViTs: 108 papers on arXiv since 2021 (data through 7 Oct 2026). 75 of its 108 mapped papers (69%) cite “Token Merging: Your ViT But Faster” (2022, arXiv 2210.09461). 50 of its 108 mapped papers (46%) cite “SPViT: Enabling Faster Vision Transformers via Soft Token Pruning” (2021, arXiv 2112.13890). 4 new papers 24 Sep–7 Oct.

What’s new

24 Sep – 7 Oct 2026

  1. This week · 2
  2. Later Is Better: Token Reduction for ViTs Under Distribution Shift
  3. Less Context, Better Geometry: Masked Geometric Encoder for Robust 3D Foundation Models also fits: Visual token pruning in multimodal models
  4. Earlier · 2
  5. DORA: Dynamic Online Reinforcement Agent for Token Pruning in Vision Transformers
  6. FlashLoop: Fast and Memory-Efficient Looped Transformers via Lazy Updates also fits: Early exit and edge inference

Foundations

Most cited by papers here

  1. Token Merging: Your ViT But Faster2022cited by 75 here
  2. SPViT: Enabling Faster Vision Transformers via Soft Token Pruning2021cited by 50 here
  3. Evo-ViT: Slow-Fast Token Evolution for Dynamic Vision Transformer2021cited by 38 here
  4. Token Fusion: Bridging the Gap between Token Pruning and Token Merging2023cited by 29 here
  5. DiffRate : Differentiable Compression Rate for Efficient Vision Transformers2023cited by 22 here

Lineage

The papers this direction builds on, oldest first.

Each paper on the line cites the one before.

  1. 2021Learned Token Pruning for Transformerscited by 17 here
  2. 2021Evo-ViT: Slow-Fast Token Evolution for Dynamic Vision Transformercited by 38 here
  3. 2021SPViT: Enabling Faster Vision Transformers via Soft Token Pruningcited by 50 here
  4. 2022Token Merging: Your ViT But Fastercited by 75 here
  5. 2023Token Merging for Fast Stable Diffusioncited by 18 here
  6. 2023Token Fusion: Bridging the Gap between Token Pruning and Token Mergingcited by 29 here
  7. 2024Accelerating Transformers with Spectrum-Preserving Token Mergingcited by 10 here
17 more it builds on, not in this chain
  1. 2022Adaptive Sparse ViT: Towards Learnable Adaptive Token Pruning by Fully Exploiting Self-Attentioncited by 10 here
  2. 2022SaiT: Sparse Vision Transformers through Adaptive Token Pruningcited by 7 here
  3. 2022HeatViT: Hardware-Efficient Adaptive Token Pruning for Vision Transformerscited by 7 here
  4. 2023Joint Token Pruning and Squeezing Towards More Aggressive Compression of Vision Transformerscited by 21 here
  5. 2023Zero-TPrune: Zero-Shot Token Pruning through Leveraging of the Attention Graph in Pre-Trained Transformerscited by 16 here
  6. 2023DiffRate : Differentiable Compression Rate for Efficient Vision Transformerscited by 22 here
  7. 2023Revisiting Token Pruning for Object Detection and Instance Segmentationcited by 15 here
  8. 2023H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Modelscited by 5 here
  9. 2023Learned Thresholds Token Merging and Pruning for Vision Transformerscited by 12 here
  10. 2023Dynamic Token Pruning in Plain Vision Transformers for Semantic Segmentationcited by 14 here
  11. 2023Which Tokens to Use? Investigating Token Reduction in Vision Transformerscited by 12 here
  12. 2023PPT: Token Pruning and Pooling for Efficient Vision Transformerscited by 10 here
  13. 2023GTP-ViT: Efficient Vision Transformers via Graph-based Token Propagationcited by 5 here
  14. 2024An Image is Worth 1/2 Tokens After Layer 2: Plug-and-Play Inference Acceleration for Large Vision-Language Modelscited by 5 here
  15. 2024ALGM: Adaptive Local-then-Global Token Merging for Efficient Semantic Segmentation with Plain Vision Transformerscited by 7 here
  16. 2024Exploring Token Pruning in Vision State Space Modelscited by 7 here
  17. 2024Learning to Merge Tokens via Decoupled Embedding for Efficient Vision Transformerscited by 8 here

The 24 most cited of 106.

Browse

30 of 108, ranked 6–35

  1. 6Joint Token Pruning and Squeezing Towards More Aggressive Compression of Vision Transformers202321
  2. 7Learned Token Pruning for Transformers202117
  3. 8Zero-TPrune: Zero-Shot Token Pruning through Leveraging of the Attention Graph in Pre-Trained Transformers202316
  4. 9Revisiting Token Pruning for Object Detection and Instance Segmentation202315
  5. 10Dynamic Token Pruning in Plain Vision Transformers for Semantic Segmentation202314
  6. 11Learned Thresholds Token Merging and Pruning for Vision Transformers202312
  7. 12Which Tokens to Use? Investigating Token Reduction in Vision Transformers202312
  8. 13Accelerating Transformers with Spectrum-Preserving Token Merging202410
  9. 14PPT: Token Pruning and Pooling for Efficient Vision Transformers202310
  10. 15Adaptive Sparse ViT: Towards Learnable Adaptive Token Pruning by Fully Exploiting Self-Attention202210
All 30
  1. 16Learning to Merge Tokens via Decoupled Embedding for Efficient Vision Transformers20248
  2. 17ALGM: Adaptive Local-then-Global Token Merging for Efficient Semantic Segmentation with Plain Vision Transformers20247
  3. 18Exploring Token Pruning in Vision State Space Models20247
  4. 19SaiT: Sparse Vision Transformers through Adaptive Token Pruning20227
  5. 20HeatViT: Hardware-Efficient Adaptive Token Pruning for Vision Transformers20227
  6. 21Token Cropr: Faster ViTs for Quite a Few Tasks20245
  7. 22GTP-ViT: Efficient Vision Transformers via Graph-based Token Propagation20235
  8. 23VLTP: Vision-Language Guided Token Pruning for Task-Oriented Segmentation20244
  9. 24Token Transforming: A Unified and Training-Free Token Compression Framework for Vision Transformer Acceleration20253
  10. 25Evict3R: Training-Free Token Eviction for Memory-Bounded Streaming Visual Geometry Transformers20253
  11. 26Faster Vision Mamba is Rebuilt in Minutes via Merged Token Re-training20243
  12. 27Dynamic Vision Mamba20252
  13. 28FastVGGT: Training-Free Acceleration of Visual Geometry Transformer20252
  14. 29LiteVGGT: Boosting Vanilla VGGT via Geometry-aware Cached Token Merging20252
  15. 30Pruning One More Token is Enough: Leveraging Latency-Workload Non-Linearities for Vision Transformers on the Edge20242
  16. 31Efficient Visual Transformer by Learnable Token Merging20242
  17. 32Vote&Mix: Plug-and-Play Token Reduction for Efficient Vision Transformer20242
  18. 33Rethinking Token Reduction for State Space Models20242
  19. 34LGViT: Dynamic Early Exiting for Accelerating Vision Transformer20232
  20. 35XStreamVGGT: Extremely Memory-Efficient Streaming Vision Geometry Grounded Transformer with KV Cache Compression20261