Hierarchical Capability Modeling
Spatial tasks are organized into three progressive levels, covering single-view perception, cross-view scene understanding, and high-level contextual reasoning.
Teaser image placeholder
Insert the paper overview figure here:
assets/teaser.png
MV-STRIDE organizes multi-view spatial reasoning into three capability levels with explicit cross-level dependencies.
Overview
Current MLLMs perform well on many single-image spatial tasks, yet still struggle to integrate evidence across viewpoints and maintain a consistent 3D understanding. MV-STRIDE addresses this bottleneck from the data side by replacing flat task collections with a hierarchical, capability-dependent organization of multi-view spatial reasoning supervision.
Capability Dependency Modeling
Cross-level task dependencies connect prerequisite perception and scene-understanding skills to complex multi-view reasoning problems.
Hierarchical Spatial Reasoning
MV-STRIDE follows a perception–modeling–reasoning hierarchy and explicitly links tasks across levels to form spatial cognition process question groups.
Single-View Spatial Perception
Build reliable egocentric spatial awareness through object location, depth, size, camera pitch, and other view-level geometric attributes.
Cross-View Scene Understanding
Bridge isolated views through object correspondence, camera motion, viewpoint transformation, and 3D object localization.
Multi-View Contextual Reasoning
Perform high-level reasoning over multiple reference frames, including counting, orientation, virtual perspective, and positional relations.
Dataset Construction Pipeline
The pipeline transforms synthetic and reconstructed 3D scenes into geometrically grounded, hierarchical multi-view QA data and long-chain reasoning supervision.
Construction pipeline placeholder
Insert the paper pipeline figure here:
assets/pipeline.png
3D-Grounded QA Generation
Extract camera and object annotations from Infinigen and ScanNet++, then generate QA pairs with verifiable geometric answers.
Cross-View Dependency Constraints
Filter questions so that the required evidence is distributed across views, reducing shortcuts that can be solved from one image.
Cognition-Guided CoT Generation
Compose prerequisite Level I and Level II sub-questions into structured reasoning supervision for Level III tasks.
Dataset at a Glance
600+Synthetic and real indoor environments
313.8KHierarchical QA pairs for Stage 1
18Spatial task categories
19.0KCoT annotations for cold start
24.8KReasoning samples for RL
Quality Verification
Sampled checks report an overall pass rate of 88.0% for QA pairs and 97.5% for CoT annotations, indicating reliable large-scale generation with limited noise.
Cross-View Dependency Validation
31.5%
Random single view
→
41.2%
Full multi-view input
A +9.7-point gain shows that the filtered Level III questions rely more strongly on cross-view evidence and reduce single-view shortcuts.
Progressive Training Strategy
MV-STRIDE supports both direct-answer spatial capability learning and structured long-chain reasoning.
Stage 1 · SFT
Data: Level I–III QA pairs with direct-answer supervision.
Goal: Establish spatial perception, cross-view understanding, and multi-view reasoning foundations.
Stage 2 · Cold Start
Data: CoT annotations synthesized from cross-level spatial cognition question groups.
Goal: Teach structured reasoning patterns and long-chain output formats.
Stage 3 · RL
Data: Difficulty-filtered multi-view reasoning problems for GRPO.
Goal: Explore and refine diverse reasoning paths while balancing accuracy and interpretability.
Main Results
Training with MV-STRIDE consistently improves spatial reasoning across single-view, multi-view, and 3D-aware benchmarks. The full-data SFT setting provides a strong accuracy-oriented baseline, while progressive training targets structured reasoning behavior.
| Model | MMSI-Bench | CV-Bench | ViewSpatial-Bench | 3DSRBench |
|---|---|---|---|---|
| Qwen3-VL-8B (Base) | 29.20 | 84.31 | 40.51 | 59.98 |
| MV-STRIDE-SFT (Full-data SFT) | 38.90 | 86.73 | 48.35 | 64.51 |
| MV-STRIDE-SFT (Stage 1) | 37.20 | 86.39 | 50.28 | 65.52 |
| MV-STRIDE-ColdStart (Stage 2) | 36.70 | 85.94 | 47.30 | 61.08 |
| MV-STRIDE-RL (Stage 3) | 37.50 | 85.94 | 49.30 | 59.43 |
Full-data SFT pools the available data for direct-answer SFT. Stage 1–3 follow the progressive QA → CoT → RL route and optimize a different objective, so benchmark accuracy is not expected to improve monotonically across stages.
Strong Spatial Reasoning Gains
Full-data SFT improves MMSI-Bench by 9.7 points over the base model and remains competitive across all four benchmarks.
Backbone-Agnostic Effectiveness
Additional experiments show consistent improvements across Qwen3-VL models of different scales and the MiMo-VL architecture.
Qualitative Reasoning
Cold-start and RL training encourage the model to produce structured reasoning that progresses from view-level perception to cross-view modeling and high-level spatial inference.
Qualitative result placeholder
Insert a representative reasoning example here:
assets/qualitative.png
BibTeX
@inproceedings{xu2026mvstride,
title = {MV-STRIDE: Enabling MLLMs to Master Multi-View Spatial Reasoning via Hierarchical Capability Modeling},
author = {Xu, Jin and Huang, Xiaojian and Luo, Zhuodong and Zhang, Zhihong and Liu, Xin and Wei, Jiansheng and Wang, Xinzhi and Zhao, Jie and Chen, Xuejin},
booktitle = {European Conference on Computer Vision},
year = {2026}
}