Research map / Vision-language and multimodal / Multimodal LLMs

Learning Multi-View Spatial Reasoning from Cross-View Relations

CVPR 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 XVR, a 100K-sample multi-view VQA dataset (general-domain WildRGB-D and robotic trajectories) on cross-view correspondence, verification, and localization, and shows fine-tuning Qwen3-VL-2B on it improves multi-view benchmarks and a GR00T-style VLA on RoboCasa.

Problem

VLMs lack multi-view spatial reasoning, producing predictions that are plausible per view but geometrically inconsistent across viewpoints; existing multi-view datasets focus on objects per view rather than inter-view geometric relations.

Approach

Generate multi-view VQA tasks grounded in SfM-inspired primitives (correspondence, verification, localization) using 3D-to-2D projection on calibrated captures and metadata-based generation on robotic trajectories, then fine-tune a VLM on the resulting data.

Key results

• Qwen3-VL-2B-XVR reaches 68.06% overall on XVR-Eval vs 36.82% for its base model (about 1.8x), above GPT-5 (61.74%) and human baseline 83.85% not exceeded overall.
• XVR training improves MindCube-Tiny Among from 32.5% to 39.5% and RoboSpatial Compatibility from 49.52% to 57.14%.
• VLA with XVR-trained backbone improves RoboCasa success by an average of 13% absolute, per the paper's abstract/introduction.

Surprising findings

• Temporal Verification accuracy declines after XVR training, the only task to drop.
• Gemini-Robotics-ER-1.5 scores 6.22% on Viewpoint Localization, below random chance (22.22%).

Limitations

• Temporal reasoning degrades after XVR training and gains are minimal on camera-motion tasks.
• VLA transfer evaluated only in simulation (RoboCasa), not on real robots.

Experimental setup

Datasets: WildRGB-D DROID RoboSet AgiBot-World FMB MobileAloha OXE MindCube RoboSpatial RoboCasa nvidia/PhysicalAI-Robotics-GR00T-X-Embodiment-Sim

Benchmarks: XVR-Eval MindCube-Tiny RoboSpatial-Home RoboCasa

Metrics: accuracy success rate

Baselines: Qwen3-VL-2B-Instruct Qwen3-VL-4B-Instruct InternVL-3.5-4B Eagle2-2B PaliGemma2-3B GPT-5 Claude-4.5-Sonnet Gemini-2.5-Pro Gemini-2.5-Flash Gemini-Robotics-ER-1.5 SpatialOM-3B SpatialMLLM-3B

Builds on: Structure-from-Motion GR00T-N1.5 Qwen3-VL MindCube RoboSpatial

Closest papers by meaning

• Seeing Across Views: Benchmarking Spatial Reasoning of Vision-Language Models in Robotic Scenes (ICLR 2026 · similarity 0.93)
• InternSpatial: A Comprehensive Dataset for Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.91)
• Seeing from Another Perspective: Evaluating Multi-View Understanding in MLLMs (AAAI 2026 · similarity 0.91)
• Spatial Understanding from Videos: Structured Prompts Meet Simulation Data (NeurIPS 2025 · similarity 0.90)
• STAR-R1: Multi-View Spatial TrAnsformation Reasoning by Reinforcing Multimodal LLMs (CVPR 2026 · similarity 0.89)
• Localizing, Structuring, and Rendering: Bridging 3D and 2D Vision-Language-Action Models for Robotic Manipulation (CVPR 2026 · similarity 0.89)