mlsanityData to 7 Oct 2026

Iterate or Widen? When Test-Time Refinement Helps LiDAR Scene Completion: A Controlled Study of Evidence Geometry, Training Coverage, and Compute

arXiv 2608.06014 ↗First posted 6 Aug 2026cs.CVcited by 0

On the map: LLM reasoning → Spatial and embodied reasoning

Abstract (arXiv)

Should a completion model spend extra test-time compute by iterating, or spend a similar parameter budget on a wider one-shot predictor? The answer is easily confounded by denoising curricula, corruption augmentation, capacity, and unpaired evaluation. We study this question in LiDAR semantic scene completion by comparing a one-shot predictor, a parameter-matched wider predictor, and a weight-tied multigrid refiner initialized from the same frozen predictor. The protocol separates coherent region removal, independent thinning, range-dependent attenuation, and additive clutter while preserving exact scene-condition pairing. Across five training seeds and 815 SemanticKITTI sequence-08 frames, the full iterative system improves mIoU over the wide control by 0.911 points under contiguous angular removal, with a 95% moving-block bootstrap interval of [0.804, 1.040] that clears a predeclared 0.5-point practical margin. Under independent 75% thinning, iteration adds only 0.300 points [0.166, 0.436], whereas observation-family augmentation adds 5.975 points [5.662, 6.140]. Neither intervention repairs additive clutter. The iterative system also costs 10.74 ms and 0.75 GiB per frame, versus 6.25 ms and 0.23 GiB for the wide control. These results establish a geometry-conditioned empirical boundary rather than a universal advantage: coherent gaps can justify fixed-depth refinement, broadly thinned evidence is addressed more effectively by training coverage, and spurious evidence requires a different robustness mechanism.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Spatial and embodied reasoning, one of 17 directions in LLM reasoning. Builds on 39 of its 47 references that are in the map. First posted on arXiv 6 Aug 2026.

Builds on

39 of 47 references.

References · 39

  1. 2019Benchmarking Neural Network Robustness to Common Corruptions and Perturbationscited by 2,857
  2. 2020RAFT: Recurrent All-Pairs Field Transforms for Optical Flowcited by 2,385
  3. 2024Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameterscited by 1,928
  4. 2022RePaint: Inpainting using Denoising Diffusion Probabilistic Modelscited by 1,419
  5. 2021Resolution-robust Large Mask Inpainting with Fourier Convolutionscited by 730
  6. 2021DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Camerascited by 673
  7. 2019Deep Equilibrium Modelscited by 603
  8. 2018Image Inpainting for Irregular Holes Using Partial Convolutionscited by 505
  9. 2016Adaptive Computation Time for Recurrent Neural Networkscited by 479
  10. 2016Semantic Scene Completion from a Single Depth Imagecited by 468
  11. 2020Measuring Robustness to Natural Distribution Shifts in Image Classificationcited by 452
  12. 2019PointRend: Image Segmentation as Renderingcited by 439
27 more
  1. 2025Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approachcited by 372
  2. 2022DiffusionDet: Diffusion Model for Object Detectioncited by 354
  3. 2023OccFormer: Dual-path Transformer for Vision-based 3D Semantic Occupancy Predictioncited by 287
  4. 2024Mixture-of-Depths: Dynamically allocating compute in transformer-based language modelscited by 242
  5. 2023Looped Transformers as Programmable Computerscited by 224
  6. 2025Reasoning with Latent Thoughts: On the Power of Looped Transformerscited by 223
  7. 2020Sparse Single Sweep LiDAR Point Cloud Segmentation via Learning Contextual Shape Priors from Scene Completioncited by 222
  8. 2020LMSCNet: Lightweight Multiscale 3D Semantic Completioncited by 187
  9. 2024SEA-RAFT: Simple, Efficient, Accurate RAFT for Optical Flowcited by 169
  10. 2025Hierarchical Reasoning Modelcited by 167
  11. 2021PonderNet: Learning to Pondercited by 156
  12. 2023Looped Transformers are Better at Learning Learning Algorithmscited by 135
  13. 2023Robo3D: Towards Robust and Reliable 3D Perception against Corruptionscited by 128
  14. 2015Human Pose Estimation with Iterative Error Feedbackcited by 121
  15. 2021Can You Learn an Algorithm? Generalizing from Easy to Hard Problems with Recurrent Networkscited by 108
  16. 2021On Interaction Between Augmentations and Corruptions in Natural Corruption Robustnesscited by 98
  17. 2023SCPNet: Semantic Scene Completion on Point Cloudcited by 80
  18. 2021Stabilizing Equilibrium Models by Jacobian Regularizationcited by 71
  19. 2023RoboBEV: Towards Robust Bird's Eye View Perception under Corruptionscited by 42
  20. 2022Path Independent Equilibrium Models Can Better Exploit Test-Time Computationcited by 33
  21. 2023SSC-RS: Elevate LiDAR Semantic Scene Completion with Representation Separation and BEV Fusioncited by 30
  22. 2024OccMamba: Semantic Occupancy Prediction with State Space Modelscited by 29
  23. 2024OccGen: Generative Multi-modal 3D Occupancy Prediction for Autonomous Drivingcited by 28
  24. 2024DiffSSC: Semantic LiDAR Scan Completion using Denoising Diffusion Probabilistic Modelscited by 16
  25. 2024TALoS: Enhancing Semantic Scene Completion via Test-time Adaptation on the Line of Sightcited by 12
  26. 2025VoxDet: Rethinking 3D Semantic Occupancy Prediction as Dense Object Detectioncited by 12
  27. 2025Enhancing 3D Semantic Scene Completion with a Refinement Modulecited by 3

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2025LiNeXt: Revisiting LiDAR Completion with Efficient Non-Diffusion Architectures
  2. 2025LiDPM: Rethinking Point Diffusion for Lidar Scene Completion
  3. 2026Generative Semantic Scene Completion
  4. 2024Distilling Diffusion Models to Efficient 3D LiDAR Scene Completion
  5. 2026PatchScene: Patch-based Voxel Diffusion for Large-Scale Scene Completion
  6. 2026ZVeC: A Zero-Shot Framework for Instance-Level Vehicle Extraction and Generative Point Cloud Completion
  7. 2025Diffusion Distillation With Direct Preference Optimization For Efficient 3D LiDAR Scene Completion
  8. 2026CloudDiffusion: Diffusion-Based Scene Completion in the Point Cloud Domain