mlsanityData to 14 Oct 2026

TTT3R: 3D Reconstruction as Test-Time Training

arXiv 2509.26645 ↗First posted 30 Sep 2025cs.CVcited by 129

Not on the map. Why?

Abstract (arXiv)

Modern Recurrent Neural Networks have become a competitive architecture for 3D reconstruction due to their linear-time complexity. However, their performance degrades significantly when applied beyond the training context length, revealing limited length generalization. In this work, we revisit the 3D reconstruction foundation models from a Test-Time Training perspective, framing their designs as an online learning problem. Building on this perspective, we leverage the alignment confidence between the memory state and incoming observations to derive a closed-form learning rate for memory updates, to balance between retaining historical information and adapting to new observations. This training-free intervention, termed TTT3R, substantially improves length generalization, achieving a 2× improvement in global pose estimation over baselines, while operating at 20 FPS with just 6 GB of GPU memory to process thousands of images. Code is available in https://rover-xingyu.github.io/TTT3R

Key facts. Cited by 129 papers in the map (data through 14 Oct 2026). Builds on 80 of its 107 references that are in the map. First posted on arXiv 30 Sep 2025.

Builds on

80 of 107 references.

References · 80

  1. 2017Attention Is All You Needcited by 60,361
  2. 2020An Image is Worth 16x16 Words: Transformers for Image Recognition at Scalecited by 30,186
  3. 2023DINOv2: Learning Robust Visual Features without Supervisioncited by 7,974
  4. 2021Emerging Properties in Self-Supervised Vision Transformerscited by 6,968
  5. 2017Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networkscited by 4,897
  6. 2016Overcoming catastrophic forgetting in neural networkscited by 4,787
  7. 2023Mamba: Linear-Time Sequence Modeling with Selective State Spacescited by 4,718
  8. 2022FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awarenesscited by 3,673
  9. 2017ScanNet: Richly-annotated 3D Reconstructions of Indoor Scenescited by 3,117
  10. 2014Sequence to Sequence Learning with Neural Networkscited by 2,986
  11. 2020Transformers are RNNs: Fast Autoregressive Transformers with Linear Attentioncited by 2,022
  12. 2021Vision Transformers for Dense Predictioncited by 1,803
38 more
  1. 2023DUSt3R: Geometric 3D Vision Made Easycited by 1,567
  2. 2025VGGT: Visual Geometry Grounded Transformercited by 1,539
  3. 2019SuperGlue: Learning Feature Matching with Graph Neural Networkscited by 1,443
  4. 2024Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Dualitycited by 1,336
  5. 2017SuperPoint: Self-Supervised Interest Point Detection and Descriptioncited by 1,320
  6. 2016Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Networkcited by 1,204
  7. 2024DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Modelcited by 1,135
  8. 2015ORB-SLAM: a Versatile and Accurate Monocular SLAM Systemcited by 988
  9. 2024Grounding Image Matching in 3D with MASt3Rcited by 983
  10. 2022Transformers learn in-context by gradient descentcited by 732
  11. 2021DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Camerascited by 687
  12. 2023Retentive Network: A Successor to Transformer for Large Language Modelscited by 613
  13. 2025Continuous 3D Perception Model with Persistent Statecited by 548
  14. 2023Gated Linear Attention Transformers with Hardware-Efficient Trainingcited by 514
  15. 2021Linear Transformers Are Secretly Fast Weight Programmerscited by 502
  16. 2016Direct Sparse Odometrycited by 485
  17. 2020Hopfield Networks is All You Needcited by 475
  18. 2024MonST3R: A Simple Approach for Estimating Geometry in the Presence of Motioncited by 442
  19. 2023Resurrecting Recurrent Neural Networks for Long Sequencescited by 408
  20. 2024Parallelizing Linear Transformers with the Delta Rule over Sequence Lengthcited by 396
  21. 2025Fast3R: Towards 3D Reconstruction of 1000+ Images in One Forward Passcited by 367
  22. 2024Learning to (Learn at Test Time): RNNs with Expressive Hidden Statescited by 355
  23. 2024Titans: Learning to Memorize at Test Timecited by 328
  24. 2024xLSTM: Extended Long Short-Term Memorycited by 327
  25. 2025Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Freecited by 283
  26. 20243D Reconstruction with Spatial Memorycited by 277
  27. 2016DSAC - Differentiable RANSAC for Camera Localizationcited by 266
  28. 2024MegaSaM: Accurate, Fast, and Robust Structure and Motion from Casual Dynamic Videoscited by 228
  29. 2023Visual Geometry Grounded Deep Structure From Motioncited by 227
  30. 2020Consistent Video Depth Estimationcited by 220
  31. 2024Repeat After Me: Transformers are Better than State Space Models at Copyingcited by 216
  32. 2024MASt3R-SLAM: Real-Time Dense SLAM with 3D Reconstruction Priorscited by 205
  33. 2025Deep Think with Confidencecited by 205
  34. 2020Robust Consistent Video Depth Estimationcited by 188
  35. 2024An Empirical Study of Mamba-based Language Modelscited by 186
  36. 2019ReFusion: 3D Reconstruction in Dynamic Environments for RGB-D Cameras Exploiting Residualscited by 181
  37. 2018BA-Net: Dense Bundle Adjustment Networkcited by 175
  38. 2025ViPE: Video Pose Engine for 3D Geometric Perceptioncited by 170

50 most cited of 80 All 80 (JSON)

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 17; 2026: 112.

129 papers.

Sort:

By area:

  1. 2026InfiniteVGGT: Visual Geometry Grounded Transformer for Endless Streamscited by 56
  2. 2026LoGeR: Long-Context Geometric Reconstruction with Hybrid Memorycited by 39
  3. 2025Human3R: Everyone Everywhere All at Oncecited by 37
  4. 2026LingBot-Map: Geometric Context Transformer for Streaming 3D Reconstructioncited by 37
  5. 2026Scal3R: Scalable Test-Time Training for Large-Scale 3D Reconstructioncited by 31
  6. 2026VGGT-SLAM 2.0: Real-time Dense Feed-forward Scene Reconstructioncited by 28
  7. 2026ZipMap: Linear-Time Stateful 3D Reconstruction via Test-Time Trainingcited by 26
  8. 2026VGGT-Omegacited by 26
  9. 2025AMB3R: Accurate Feed-forward Metric-scale 3D Reconstruction with Backendcited by 20
  10. 2025RealX3D: A Physically-Degraded 3D Benchmark for Multi-view Visual Restoration and Reconstructioncited by 19
  11. 2026RoboTTT: Context Scaling for Robot Policiescited by 19
  12. 2025ViT3: Unlocking Test-Time Training in Visioncited by 18
38 more
  1. 2026VGG-T3: Offline Feed-Forward 3D Reconstruction at Scalecited by 18
  2. 2025FlashVGGT: Efficient and Scalable Visual Geometry Transformers with Compressed Descriptor Attentioncited by 17
  3. 2025TeleWorld: Towards Dynamic Multimodal Synthesis with a 4D World Modelcited by 16
  4. 2026Motion 3-to-4: 3D Motion Reconstruction for 4D Synthesiscited by 15
  5. 2026AnchorWeave: World-Consistent Video Generation with Retrieved Local Spatial Memoriescited by 15
  6. 2026TokenGS: Decoupling 3D Gaussian Prediction from Pixels with Learnable Tokenscited by 15
  7. 2026Test-Time Training with KV Binding Is Secretly Linear Attentioncited by 13
  8. 2026Track4World: Feedforward World-centric Dense 3D Tracking of All Pixelscited by 13
  9. 2026tttLRM: Test-Time Training for Long Context and Autoregressive 3D Reconstructioncited by 12
  10. 2026GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokenscited by 12
  11. 2026TTSA3R: Training-Free Temporal-Spatial Adaptive Persistent State for Streaming 3D Reconstructioncited by 11
  12. 2026DROID-SLAM in the Wildcited by 10
  13. 2025HTTM: Head-wise Temporal Token Merging for Faster VGGTcited by 9
  14. 2026HorizonStream: Long-Horizon Attention for Streaming 3D Reconstructioncited by 9
  15. 2026VGGT-Motion: Motion-Aware Calibration-Free Monocular SLAM for Long-Range Consistencycited by 8
  16. 2026MeMix: Writing Less, Remembering More for Streaming 3D Reconstructioncited by 8
  17. 2026Vision Pretraining for Dense Spatial Perceptioncited by 8
  18. 2025MUT3R: Motion-aware Updating Transformer for Dynamic 3D Reconstructioncited by 7
  19. 2025SplatPainter: Interactive Authoring of 3D Gaussians from 2D Edits via Test-Time Trainingcited by 7
  20. 2026XStreamVGGT: Extremely Memory-Efficient Streaming Vision Geometry Grounded Transformer with KV Cache Compressioncited by 7
  21. 2026Stream3D: Sequential Multi-View 3D Generation via Evidential Memorycited by 7
  22. 2025VIGS-SLAM: Visual Inertial Gaussian Splatting SLAMcited by 6
  23. 2025LASER: Layer-wise Scale Alignment for Training-Free Streaming 4D Reconstructioncited by 6
  24. 2025WorldWarp: Propagating 3D Geometry with Asynchronous Video Diffusioncited by 6
  25. 2026Mem3R: Streaming 3D Reconstruction with Hybrid Memory via Test-Time Trainingcited by 6
  26. 2026Feed-Forward 3D Scene Modeling: A Problem-Driven Perspectivecited by 6
  27. 2025KV-Tracker: Real-Time Pose Tracking with Transformerscited by 5
  28. 2026Action-Geometry Prediction with 3D Geometric Prior for Bimanual Manipulationcited by 5
  29. 2026StreamCacheVGGT: Streaming Visual Geometry Transformers with Robust Scoring and Hybrid Cache Compressioncited by 5
  30. 2025SaLon3R: Structure-aware Long-term Generalizable 3D Reconstruction from Unposed Imagescited by 4
  31. 2025DriveVGGT: Calibration-Constrained Visual Geometry Transformers for Multi-Camera Autonomous Drivingcited by 4
  32. 2025MoonSeg3R: Monocular Online Zero-Shot Segment Anything in 3D with Reconstructive Foundation Priorscited by 4
  33. 2026Mix3R: Mixing Feed-forward Reconstruction and Generative 3D Priors for Joint Multi-view Aligned 3D Reconstruction and Pose Estimationcited by 4
  34. 2026MERG3R: A Divide-and-Conquer Approach to Large-Scale Neural Visual Geometrycited by 3
  35. 2026PROSPECT: Unified Streaming Vision-Language Navigation via Semantic--Spatial Fusion and Latent Predictive Representationcited by 3
  36. 2026FrameVGGT: Coherence-Preserving Memory for Bounded Streaming Geometrycited by 3
  37. 2026FILT3R: Latent State Adaptive Kalman Filter for Streaming 3D Reconstructioncited by 3
  38. 2026RayMap3R: Inference-Time RayMap for Dynamic 3D Reconstructioncited by 3

50 most cited of 129 All 129 (JSON)