mlsanityData to 14 Oct 2026

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

arXiv 2509.17650 ↗First posted 22 Sep 2025cs.CVcited by 14

On the map: Efficient LLM inference and serving → Token pruning and merging for ViTs

Abstract (arXiv)

Streaming visual transformers like StreamVGGT achieve strong 3D perception but suffer from unbounded growth of key value (KV) memory, which limits scalability. We propose a training-free, inference-time token eviction policy that bounds memory by discarding redundant tokens while keeping the most informative ones. Our method uses significantly less memory with little to no drop in accuracy: on 7-Scenes with long sequences it reduces peak memory from 18.63 GB to 9.39 GB while accuracy and completeness drop by only 0.003. Under strict memory budgets, eviction enables denser frame sampling, which improves reconstruction accuracy compared to the baseline. Experiments across video depth estimation (Sintel, KITTI), 3D reconstruction (7-Scenes, NRGBD), and camera pose estimation (Sintel, TUM-dynamics) show that our approach closely matches StreamVGGT at a fraction of the memory and makes long-horizon streaming inference more practical.

Key facts. Cited by 14 papers in the map (data through 14 Oct 2026). Placed in Token pruning and merging for ViTs, one of 12 directions in Efficient LLM inference and serving. Builds on 18 of its 23 references that are in the map. First posted on arXiv 22 Sep 2025.

Builds on

18 of 23 references.

Closest ancestors · 1

Same direction, newest first

  1. 2022Token Merging: Your ViT But Faster17 Oct

Other references · 17

  1. 2023DINOv2: Learning Robust Visual Features without Supervisioncited by 7,974
  2. 2021Vision Transformers for Dense Predictioncited by 1,803
  3. 2023DUSt3R: Geometric 3D Vision Made Easycited by 1,567
  4. 2025VGGT: Visual Geometry Grounded Transformercited by 1,539
  5. 2024Grounding Image Matching in 3D with MASt3Rcited by 983
  6. 2021DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsificationcited by 864
  7. 2023H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Modelscited by 832
  8. 2025Continuous 3D Perception Model with Persistent Statecited by 548
  9. 2024MonST3R: A Simple Approach for Estimating Geometry in the Presence of Motioncited by 442
  10. 2023Scissorhands: Exploiting the Persistence of Importance Hypothesis for LLM KV Cache Compression at Test Timecited by 439
  11. 2024PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funnelingcited by 432
  12. 2025Fast3R: Towards 3D Reconstruction of 1000+ Images in One Forward Passcited by 367
5 more
  1. 20243D Reconstruction with Spatial Memorycited by 277
  2. 2025Streaming 4D Visual Geometry Transformercited by 161
  3. 2025Point3R: Streaming 3D Reconstruction with Explicit Spatial Pointer Memorycited by 87
  4. 2023PuMer: Pruning and Merging Tokens for Efficient Vision Language Modelscited by 53
  5. 2024D2O: Dynamic Discriminative Operations for Efficient Long-Context Inference of Large Language Modelscited by 31

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 2; 2026: 12.

14 papers.

Sort:

By area:

  1. 2026VGGT-Omegacited by 26
  2. 2025AMB3R: Accurate Feed-forward Metric-scale 3D Reconstruction with Backendcited by 20
  3. 2026MeMix: Writing Less, Remembering More for Streaming 3D Reconstructioncited by 8
  4. 2026Feed-Forward 3D Scene Modeling: A Problem-Driven Perspectivecited by 6
  5. 2026OnlineHMR: Video-based Online World-Grounded Human Mesh Recoverycited by 4
  6. 2025On-the-fly Large-scale 3D Reconstruction from Multi-Camera Rigscited by 3
  7. 2026FrameVGGT: Coherence-Preserving Memory for Bounded Streaming Geometrycited by 3
  8. 2026GHOST: Geometry-Hierarchical Online Streaming Token Eviction for Efficient 3D Reconstructioncited by 2
  9. 2026OVGGT: O(1) Constant-Cost Streaming Visual Geometry Transformercited by 1
  10. 2026Good Token Hunting: A Hitchhiker's Guide to Token Selection for Visual Geometry Transformerscited by 0
  11. 2026VI3: Grounding Pretrained 3D Foundation Models with Inertial Cuescited by 0
  12. 2026Scal3R: Learning Efficient Multi-Relative Pose Query for Scalable Online 3D Reconstructioncited by 0
2 more
  1. 2026AMB3R-SLAM: Kilometer-scale SLAM with Hierarchical Backendcited by 0
  2. 2026RegVGGT: Sustainable Visual Geometry Grounding for Streaming via Regulated Memorycited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026XStreamVGGT: Extremely Memory-Efficient Streaming Vision Geometry Grounded Transformer with KV Cache Compression
  2. 2026XStreamVGGT: Extremely Memory-Efficient Streaming Vision Geometry Grounded Transformer with KV Cache Compression
  3. 2026InfiniteVGGT: Visual Geometry Grounded Transformer for Endless Streams
  4. 2026Stream3D: Sequential Multi-View 3D Generation via Evidential Memory
  5. 2026LongStream: Long-Sequence Streaming Autoregressive Visual Geometry
  6. 2026Spark3R: Asymmetric Token Reduction Makes Fast Feed-Forward 3D Reconstruction
  7. 2025FastVGGT: Training-Free Acceleration of Visual Geometry Transformer
  8. 2025LiteVGGT: Boosting Vanilla VGGT via Geometry-aware Cached Token Merging