Research map / Vision-language and multimodal / Multimodal LLMs
Seeing Across Views: Benchmarking Spatial Reasoning of Vision-Language Models in Robotic Scenes
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 MV-RoboBench, a 1.7k-item manually curated multi-view multiple-choice benchmark for spatial and robotic manipulation reasoning, and evaluates many VLMs plus CoT-style enhancements, finding a large gap to humans.
Problem
Whether VLMs can integrate multi-view camera inputs for spatial understanding and robotic manipulation decisions remains unevaluated, since existing spatial benchmarks are mostly single-view or not embodied.
Approach
A manually curated multi-view QA benchmark built from real robot demonstrations (AgiWorld, BridgeV2) spanning four spatial and four robotic-execution subtasks, plus analysis of correlation between spatial and robotic accuracy and transfer from single-view OmniSpatial results.
Key results
- GPT-5 achieves the highest overall accuracy on MV-RoboBench at 56.41%, versus 91.04% for human participants.
- Most non-reasoning open-source VLMs perform near random choice (19.71%), with Qwen2.5-VL-72B best among open-source at 24.29%.
- Depth priors give GPT-4.1 a +3.25 point average gain (29.87 to 33.12), while synthetic VGGT views generally degrade performance.
Surprising findings
- Strong performance on single-view OmniSpatial does not reliably transfer to multi-view robotic reasoning for most models.
- Synthetic novel views (VGGT) more often hurt performance than help, and CoT effects are mixed and model-dependent.
Limitations
- Human baseline uses only five computer-science participants and no time limit.
- Some benchmark items rely on GPT-4.1 for candidate filtering and text polishing.
- Novel view synthesis methods tested performed poorly, limiting the visual CoT analysis.
Experimental setup
Datasets: AgiWorld BridgeV2 OmniSpatial
Benchmarks: MV-RoboBench OmniSpatial ERQA
Metrics: accuracy
Baselines: GPT-4o GPT-4.1 GPT-5 Gemini-2.5-pro o4-mini Qwen2.5-VL-72B InternVL3-78B Llama-4-Maverick Gemma-3-27B Claude-3.7
Builds on: Chain-of-Thought prompting VGGT MoGe-2 OmniSpatial
Closest papers by meaning
- Learning Multi-View Spatial Reasoning from Cross-View Relations (CVPR 2026 · similarity 0.93)
- Seeing from Another Perspective: Evaluating Multi-View Understanding in MLLMs (AAAI 2026 · similarity 0.93)
- Multi-SpatialMLLM: Multi-Frame Spatial Understanding with Multi-Modal Large Language Models (CVPR 2026 · similarity 0.93)
- InternSpatial: A Comprehensive Dataset for Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.92)
- Abstract 3D Perception for Spatial Intelligence in Vision-Language Models (CVPR 2026 · similarity 0.91)
- PIVOT: Iterative Visual Prompting Elicits Actionable Knowledge for VLMs (ICML 2024 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.