Research map / Vision-language and multimodal / Multimodal LLMs

SpaceVista: All-Scale Visual Spatial Reasoning from mm to km

ICML 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.

Builds SpaceVista-1M, a 1M-QA all-scale (mm to km) spatial video reasoning dataset from 38K scenes across 19 tasks, a hand-curated SpaceVista-Bench, and SpaceVista-7B (Qwen2.5-VL-7B with DINOv3 fusion, LoRA-like scale experts, and anchor-based progressive RL rewards).

Problem

Spatial reasoning MLLMs are largely limited to indoor scenes and constrained scales, with no training data or framework covering objects and scenes from millimeters to kilometers, and no reliable all-scale evaluation.

Approach

An automated specialist-driven pipeline (metric depth, video depth, grounding, SAM 2) builds multi-scale spatial QA from existing 3D-annotated videos; the model fuses DINOv3 dense features via cross-attention, uses scale-specific LoRA-like experts with a learned router to reduce cross-scale knowledge conflict, and applies RL with progressive scale and semantic rewards gated behind answer correctne

Key results

• SpaceVista-7B with RL reaches 39.2 on SpaceVista-Bench overall, versus 33.8 for Gemini-2.5-pro and 33.7 for GPT-5 (the best closed models listed).
• SpaceVista-7B w/ RL scores 48.6 on VSI-Bench and 41.6 on SPAR-Bench, improving over Qwen2.5-VL-7B (32.7 and 33.1).
• On the 3B ablation, the scale expert with M=4 gives +2.1 on VSI-Bench and +3.7 on SpaceVista-Bench over vanilla SFT.

Surprising findings

• Increasing the number of scale experts beyond 4 (M=6) hurts performance, attributed to router load imbalance.
• DINOv3 dense features yielded larger gains than VGGT on the benchmarks, despite VGGT being geometry-focused.

Limitations

• Drone-scene data is small and exploration there is described as preliminary.
• Automated labels from specialist models are not fully reliable; manual QA check on training set reports about 83% accuracy.
• Evaluation on SpaceVista-Bench uses only about 1,600 QA pairs over roughly 300 scenes.

Experimental setup

Datasets: SpaceVista-1M uCO3D WildRGB-D SMOT ScanNet ScanNet++ DL3DV SpaceR SPAR-Bench

Benchmarks: VSI-Bench STI-Bench SpaceVista-Bench MMSI-Bench SPAR-Bench

Metrics: accuracy

Baselines: GPT-5 Gemini-2.5-pro InternVL3.5-38B Qwen2.5-VL-7B VLM-3R Spatial-MLLM SpaceR VILASR VG LLM

Builds on: GRPO LoRA DINOv3 SAM 2 Qwen2.5-VL

Closest papers by meaning

• Thinking in Structures: Evaluating Spatial Intelligence in Constraint-Governed Spaces (ICML 2026 · similarity 0.91)
• Spatial-DISE: A Unified Benchmark for Evaluating Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.91)
• SpatiaLab: Can Vision–Language Models Perform Spatial Reasoning in the Wild? (ICLR 2026 · similarity 0.90)
• From Indoor to Open World: Revealing the Spatial Reasoning Gap in MLLMs (CVPR 2026 · similarity 0.90)
• InternSpatial: A Comprehensive Dataset for Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.90)
• SpatialRGPT: Grounded Spatial Reasoning in Vision-Language Models (NeurIPS 2024 · similarity 0.90)