Uncovering and Shaping the Latent Representation of 3D Scene Topology in Vision-Language Models

• URL: http://arxiv.org/abs/2605.07148v1
• Date: Fri, 08 May 2026 02:32:27 GMT
• Title: Uncovering and Shaping the Latent Representation of 3D Scene Topology in Vision-Language Models
• Abstract summary: We show that modern Vision-Language Models (VLMs) possess a latent topological map of 3D scenes.<n>By isolating this spatial subspace through cross-scene linear feature extraction, we extract a clean spatial subspace that causally controls the model's spatial outputs.<n>We mathematically shape this latent representation and prove its correspondence to the Laplacian eigenmaps of the scene's 3D Gaussian- kernel graph.
• Score: 9.722829662835233
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Decades of cognitive science establish that humans navigate environments by forming cognitive maps, defined as allocentric and topology-preserving representations of 3D space. While modern Vision-Language Models (VLMs) demonstrate emergent spatial reasoning from 2D egocentric inputs, it remains unclear whether they construct an analogous 3D internal representation. In this paper, we demonstrate that current VLMs do possess a latent topological map of 3D scenes, but it is heavily overshadowed by non-geometric visual semantics, such as color and shape. By isolating this spatial subspace through cross-scene linear feature extraction, we extract a clean spatial subspace that causally controls the model's spatial outputs. We mathematically shape this latent representation and prove its correspondence to the Laplacian eigenmaps of the scene's 3D Gaussian-kernel graph, converging to the physical 3D space in the continuous limit. Motivated by this geometric identification, we further introduce a mathematically principled latent regularization method for VLMs, based on Dirichlet energy. Applying this single-term regularizer to a minimal 500-step supervised VLM fine-tuning (SFT) on simple synthetic data yields significant improvements on real-world spatial benchmarks, outperforming standard SFT and competitive baselines by up to 12.1\% in spatial tasks involving scene topology understanding. Source code is available at https://github.com/pittisl/vlm-latent-shaping

Related papers

• Sparse auto-regressive modeling for scene generation from multi-view images [20.272184280044083]
  We introduce SPAR3S, a sparse voxel-aligned 3D latent generative model for conditional scene completion.<n>Our key insight is to formulate 3D scene generation in a structured, compact, voxel-aligned 3D latent space.<n>We demonstrate the effectiveness of our method on synthetic indoor scenes, achieving higher novel-view quality than prior work.
  arXiv  Detail & Related papers  (2026-09-03T14:41:34Z)
• SeeSE3: Emergence of 3D Space in Vision Features [54.140615953727355]
  We investigate the relation between the structure of the space of visual features and the group of Euclidean transformations $SE(3)$.<n>We show that self-supervised vision models, which, in principle, have not been trained with direct 3D supervision or active agency, possess latent subspaces that are remarkably strongly correlated with 3D Euclidean space, when probed correctly.<n>We propose a new class of "Latent-Space Navigation" techniques that perform visual odometry and localization purely in the latent space, bypassing the need for explicit 3D reconstruction.
  arXiv  Detail & Related papers  (2026-07-15T18:00:31Z)
• VEGA: Visual Encoder Grounding Alignment for Spatially-Aware Vision-Language-Action Models [55.12929235609365]
  Visual backbones of current vision-language-action (VLA) models are predominantly pretrained on 2D image data without explicit 3D geometric supervision.<n>Existing implicit grounding methods partially address this by aligning VLA features with those of 3D-aware foundation models.<n>We propose VEGA, a framework that directly aligns the output of the VLA's visual encoder with spatially-aware features from DINOv2-FiT3D.
  arXiv  Detail & Related papers  (2026-05-11T12:44:26Z)
• SpaceMind++: Toward Allocentric Cognitive Maps for Spatially Grounded Video MLLMs [6.871639227720094]
  We propose SpaceMind, a video MLLM architecture that builds a voxelized cognitive map from RGB videos.<n>This map reorganizes fragmented egocentric observations into a shared 3D metric representation.<n>We introduce Coordinate Deep Iterative Fusion, a new mechanism that relays map-level spatial knowledge back into the original 2D visual features.
  arXiv  Detail & Related papers  (2026-05-10T10:01:57Z)
• Proxy3D: Efficient 3D Representations for Vision-Language Models via Semantic Clustering and Alignment [50.02189698630855]
  We propose a Proxy3D method with compact yet comprehensive 3D proxy representations for the vision modality.<n>Our method achieves competitive or state-of-the-art performance in 3D visual question answering, visual grounding and general spatial intelligence benchmarks.
  arXiv  Detail & Related papers  (2026-05-08T17:50:47Z)
• Cog3DMap: Multi-View Vision-Language Reasoning with 3D Cognitive Maps [77.63233146945718]
  We introduce Cog3DMap, a framework that recurrently constructs an explicit 3D memory from multi-view images.<n>Our framework enables direct reasoning over a spatially structured 3D map, achieving state-of-the-art performance on various spatial reasoning benchmarks.
  arXiv  Detail & Related papers  (2026-03-24T10:05:32Z)
• Generation Models Know Space: Unleashing Implicit 3D Priors for Scene Understanding [50.098085774845195]
  We propose a paradigm shift by leveraging the implicit spatial prior within large-scale generation models.<n>We introduce GeneEGA-3D (Video Extracted Generative Awareness), a plug-and-play framework that repurposes a pre-trained video diffusion model as a Latent World Simulator.
  arXiv  Detail & Related papers  (2026-03-19T17:59:58Z)
• OneWorld: Taming Scene Generation with 3D Unified Representation Autoencoder [90.8453349494245]
  We present OneWorld, a framework that performs diffusion directly within a coherent 3D representation space.<n>OneWorld generates high-quality 3D scenes with superior cross-view consistency compared to state-of-the-art 2D-based methods.
  arXiv  Detail & Related papers  (2026-03-17T03:43:37Z)
• Spa3R: Predictive Spatial Field Modeling for 3D Visual Reasoning [43.746951848993035]
  spatial intelligence can emerge from 2D vision alone, rather than being imposed via explicit spatial instruction tuning.<n>We introduce Spa3R, a self-supervised framework that learns a unified view-in spatial representation directly from unposed multi-view images.<n>Experiments demonstrate Spa3-VLM achieves state-of-the-art accuracy of 58.6% on 3D VQA, significantly outperforming prior methods.
  arXiv  Detail & Related papers  (2026-02-24T18:37:34Z)
• Cross-Modal Geometric Hierarchy Fusion: An Implicit-Submap Driven Framework for Resilient 3D Place Recognition [9.411542547451193]
  We propose a novel framework that redefines 3D place recognition through density-agnostic geometric reasoning.<n>Specifically, we introduce an implicit 3D representation based on elastic points, which is immune to the interference of original scene point cloud density.<n>With the aid of these two types of information, we obtain descriptors that fuse geometric information from both bird's-eye view and 3D segment perspectives.
  arXiv  Detail & Related papers  (2025-06-17T07:04:07Z)
• Flattening-Net: Deep Regular 2D Representation for 3D Point Cloud Analysis [66.49788145564004]
  We present an unsupervised deep neural architecture called Flattening-Net to represent irregular 3D point clouds of arbitrary geometry and topology. Our methods perform favorably against the current state-of-the-art competitors.
  arXiv  Detail & Related papers  (2022-12-17T15:05:25Z)
• Cylinder3D: An Effective 3D Framework for Driving-scene LiDAR Semantic Segmentation [87.54570024320354]
  State-of-the-art methods for large-scale driving-scene LiDAR semantic segmentation often project and process the point clouds in the 2D space. A straightforward solution to tackle the issue of 3D-to-2D projection is to keep the 3D representation and process the points in the 3D space. We develop a 3D cylinder partition and a 3D cylinder convolution based framework, termed as Cylinder3D, which exploits the 3D topology relations and structures of driving-scene point clouds.
  arXiv  Detail & Related papers  (2020-08-04T13:56:19Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.