Tsukada Lab. Computer Network Group

Evict3R: Training-Free Token Eviction for Memory-Bounded Streaming Visual Geometry Transformers

Soroush Mahdi, Fardin Ayar, Ehsan Javanmardi, Manabu Tsukada, Mahdi Javanmardi

, Sep 22, 2025

Summary in three lines

  • Evict3R limits the key–value memory of streaming visual geometry transformers by evicting redundant tokens at inference.
  • It runs without additional training and leaves room to sample more frames under tight memory limits.
  • On long 7-Scenes sequences, peak memory fell from 18.63 GB to 9.39 GB with only a small reconstruction-quality change.

Figures from the paper

Figure 2: Evict3R token eviction framework
Figure 2. Evict3R token eviction frameworkSource: Figure 2 from Mahdi et al., arXiv (2025) (authors’ arXiv version 2509.17650, CC BY 4.0)The diagram shows how the policy scores tokens and limits the cache separately for each layer.

How to cite

S. Mahdi, F. Ayar, E. Javanmardi, M. Tsukada, and M. Javanmardi, “Evict3R: Training-Free Token Eviction for Memory-Bounded Streaming Visual Geometry Transformers,” in , 2025.

BibTeX

@unpublished{Mahdi2025,
  title         = {Evict3R: Training-Free Token Eviction for Memory-Bounded Streaming Visual Geometry Transformers},
  author        = {Soroush Mahdi and Fardin Ayar and Ehsan Javanmardi and Manabu Tsukada and Mahdi Javanmardi},
  year          = {2025},
  date          = {2025-09-22},
  eprint        = {2509.17650},
  archiveprefix = {arXiv},
  primaryclass  = {cs.CV},
  url           = {https://arxiv.org/abs/2509.17650},
  note          = {Preprint}
}