Research map / Vision-language and multimodal / Multimodal LLMs

OmniSpatial: Towards Comprehensive Spatial Reasoning Benchmark for Vision Language Models

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs

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 OmniSpatial, a 8.4K-question spatial reasoning benchmark across four cognitive categories (50 subtasks), evaluates many VLMs, and proposes PointGraph and SpatialCoT prompting strategies that give modest gains.

Problem

Existing spatial reasoning benchmarks focus on basic relations (left/right, near/far, counting) that recent reasoning VLMs have nearly saturated, leaving complex spatial reasoning under-evaluated.

Approach

A cognitively grounded taxonomy of four spatial reasoning dimensions (dynamic reasoning, complex spatial logic, spatial interaction, perspective-taking) with manually annotated, diverse, non-template questions from web images, exams, driving tests and existing datasets; plus two training-free augmentations: PointGraph (Florence-2 scene graph in prompt) and SpatialCoT (InstantMesh novel views colla

Key results

• Best model o3 reaches 56.33% average accuracy on OmniSpatial-test versus 92.63% for humans.
• PointGraph raises GPT-4.1-mini from 48.86 to 50.49 and Gemini-2.5-Flash from 51.47 to 53.23 average accuracy.
• SFT of Qwen-VL2.5-3B on 6.9K OmniSpatial-train improves average from 40.30 to 48.12, versus +1.29 from a 200K template corpus.

Surprising findings

• Traditional textual chain-of-thought gives only marginal gains (under +1 point) on OmniSpatial.
• Blind (no-image) GPT-3.5/GPT-4 baselines score near random on some tracks, yet a 200K template corpus adds little, suggesting template data is less useful than manually curated data.

Limitations

• Benchmark questions are primarily static or sparsely sampled frames, so dynamic reasoning is only partially tested.
• Gains from PointGraph and SpatialCoT are modest (about 1-2 points) and inconsistent across tracks.
• Multiple-choice format and reliance on LLM-as-judge for reasoning models may introduce parsing bias.

Experimental setup

Datasets: OmniSpatial-train MME HOI4D SpatialViz PhysBench ViewSpatial DrivingVQA VSI-Bench SpaceR

Benchmarks: OmniSpatial-test VSI-Bench SpatialBot-Bench EmbSpatial

Metrics: accuracy

Baselines: GPT-4o GPT-4.1 Gemini-2.5-pro InternVL3-78B Qwen-VL2.5-72B SoFar RoboPoint SpatialBot SpaceThinker

Builds on: SpaceR VSI-Bench SpatialVLM InstantMesh Florence-2

Closest papers by meaning

• Spatial-DISE: A Unified Benchmark for Evaluating Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.92)
• Is A Picture Worth A Thousand Words? Delving Into Spatial Reasoning for Vision Language Models (NeurIPS 2024 · similarity 0.92)
• SpatiaLab: Can Vision–Language Models Perform Spatial Reasoning in the Wild? (ICLR 2026 · similarity 0.92)
• InternSpatial: A Comprehensive Dataset for Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.91)
• SpatiaLQA: A Benchmark for Evaluating Spatial Logical Reasoning in Vision-Language Models (CVPR 2026 · similarity 0.91)
• SpatialRGPT: Grounded Spatial Reasoning in Vision-Language Models (NeurIPS 2024 · similarity 0.90)