Training Noise Token Pruning
On the map: Efficient LLM inference and serving → Token pruning and merging for ViTs
Abstract (arXiv)
In the present work we present Training Noise Token (TNT) Pruning for vision transformers. Our method relaxes the discrete token dropping condition to continuous additive noise, providing smooth optimization in training, while retaining discrete dropping computational gains in deployment settings. We provide theoretical connections to Rate-Distortion literature, and empirical evaluations on the ImageNet dataset using ViT and DeiT architectures demonstrating TNT's advantages over previous pruning methods.
Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Token pruning and merging for ViTs, one of 12 directions in Efficient LLM inference and serving. Builds on 16 of its 23 references that are in the map. First posted on arXiv 27 Nov 2024.
Builds on
16 of 23 references.
Closest ancestors · 5
Same direction, newest first
- 2023Which Tokens to Use? Investigating Token Reduction in Vision Transformers9 Aug
- 2023Zero-TPrune: Zero-Shot Token Pruning through Leveraging of the Attention Graph in Pre-Trained Transformers27 May
- 2022Token Merging: Your ViT But Faster17 Oct
- 2021Evo-ViT: Slow-Fast Token Evolution for Dynamic Vision Transformer3 Aug
- 2021Learned Token Pruning for Transformers2 Jul
Other references · 11
- 2017Attention Is All You Needcited by 59,932
- 2020An Image is Worth 16x16 Words: Transformers for Image Recognition at Scalecited by 29,915
- 2020Training data-efficient image transformers & distillation through attentioncited by 4,560
- 2016Deep Variational Information Bottleneckcited by 915
- 2021DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsificationcited by 849
- 2021AdaViT: Adaptive Tokens for Efficient Vision Transformercited by 360
- 2022Not All Patches are What You Need: Expediting Vision Transformers via Token Reorganizationscited by 359
- 2023Token Merging for Fast Stable Diffusioncited by 220
- 2021Adaptive Token Sampling For Efficient Vision Transformerscited by 181
- 2016Information Dropout: Learning Optimal Representations Through Noisy Computationcited by 131
- 2020Length-Adaptive Transformer: Train Once with Length Drop, Use Anytime with Searchcited by 76
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2023Learned Thresholds Token Merging and Pruning for Vision Transformers
- 2025Similarity-Aware Token Pruning: Your VLM but Faster
- 2023PPT: Token Pruning and Pooling for Efficient Vision Transformers
- 2025Token Pruning in Audio Transformers: Optimizing Performance and Decoding Patch Importance
- 2023Joint Token Pruning and Squeezing Towards More Aggressive Compression of Vision Transformers
- 2021SPViT: Enabling Faster Vision Transformers via Soft Token Pruning
- 2026DORA: Dynamic Online Reinforcement Agent for Token Pruning in Vision Transformers
- 2023DiffRate : Differentiable Compression Rate for Efficient Vision Transformers