Research map / Vision-language and multimodal / Multimodal LLMs

SpatialBench: Is Your Spatial Foundation Model an All-Round Player

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces SpatialBench, a deterministic multi-density, cross-paradigm benchmark (19 datasets, 546 scenes, 41 spatial foundation model variants), plus DA-Next and the DA-Next-5M egocentric/wrist-view dataset to close the embodied domain gap.

Problem

Existing evaluations of spatial foundation models cover narrow paradigms, use non-standardized scene splits and frame sampling, and rarely test scaling with input density or limited domains such as egocentric and wrist-view data.

Approach

A deterministic benchmark that evaluates each scene at four view-density regimes (single, sparse via set-cover, medium, dense) across six reconstruction paradigms and five domain-tagged task suites, paired with a wrist/egocentric training dataset and a Depth-Anything-3-based model with scale tokens.

Key results

• DA-Next reduces sparse-view depth AbsRel from 0.095 (DA3-Giant) to 0.050 (-47.4%) and medium-view from 0.086 to 0.035 (-59.3%).
• DA-Next improves AUC@30 by +3.1% (sparse) and +5.5% (medium) over DA3-Giant.
• On the dense regime, TTT3R nearly doubles AUC@30 over CUT3R (0.165 to 0.321) and LoGeR lifts Pi3's dense AUC@30 by 14.1%.

Surprising findings

• Vanilla VGGT runs out of memory under dense inputs, and TTT methods like TTT3R regress on sparse inputs versus their base models.
• Injecting GT camera poses does not uniformly help: OmniVGGT, Pi3-X and WorldMirror partially override the injected poses.

Limitations

• DA-Next does not run on the dense regime (OOM), and its ATE is worse than DA3-Giant on medium inputs (1.442 vs 1.161).
• Many full-context models OOM or time out under dense inputs, so dense comparisons are partial.
• Dense-regime per-dataset results are missing for some datasets, since they are not eligible for dense evaluation.

Experimental setup

Datasets: DA-Next-5M Mapfree Hypersim Infinigen DROID Xperience RLBench RoboTwin RoboLab ADT 7-Scenes DTU

Benchmarks: SpatialBench

Metrics: AbsRel AUC@30 ATE F-Score RPE Accuracy Completeness

Baselines: DA3-Giant VGGT Pi3-X MapAnything OmniVGGT WorldMirror CUT3R StreamVGGT LoGeR Scal3R

Builds on: Depth Anything 3 VGGT DUSt3R MapAnything TTT3R

Closest papers by meaning

• Scaling Spatial Intelligence with Multimodal Foundation Models (CVPR 2026 · similarity 0.89)
• SpaceVista: All-Scale Visual Spatial Reasoning from mm to km (ICML 2026 · similarity 0.88)
• Spatial457: A Diagnostic Benchmark for 6D Spatial Reasoning of Large Mutimodal Models (CVPR 2025 · similarity 0.87)
• Exploring Spatial Intelligence from a Generative Perspective (CVPR 2026 · similarity 0.87)
• MMSI-Bench: A Benchmark for Multi-Image Spatial Intelligence (ICLR 2026 · similarity 0.87)
• InfiniBench: Infinite Benchmarking for Visual Spatial Reasoning with Customizable Scene Complexity (CVPR 2026 · similarity 0.86)