ConsiSpace: Learning Geometric Consistency Matters for Video Spatial Reasoning
ConsiSpace leverages geometric consistency to enhance video spatial reasoning, improving stability and efficiency with a 12.6-point gain.
Key Findings
Methodology
ConsiSpace employs geometric consistency as a core principle, constructing a Geometry-Consistent Memory (GCM) that stores implicit evidence tokens and explicit geometric cues. It uses geometry-guided writing, fusion, and hierarchical retrieval to organize evidence efficiently. The model integrates a unified self-supervised reinforcement learning (UC-SSRL) framework, which applies reward signals based on answer, metric, and topological consistency across multiple views. The process involves: • extracting visual features via SigLIP2; • encoding geometric cues with VGGT; • gating evidence writing based on inter-frame geometric changes; • fusing spatially similar evidence within geometric neighborhoods; • retrieving relevant evidence using multi-view sampling; • constructing context tokens for the language model; • applying UC-SSRL to refine cross-view stability.
Key Results
- On VSI-Bench, OSI-Bench, and MMSI-Video-Bench, ConsiSpace achieved an average score increase of 12.6 points, with a maximum of 79.9, outperforming strong baselines. It demonstrated significant improvements in spatial relation, measurement, and path planning tasks, with accuracy gains exceeding 15% in some cases.
- In VSI-Bench, the average accuracy reached 73.8%, surpassing previous bests by 12.6 points. The model excelled particularly in spatial relation and measurement tasks. On MMSI-Video-Bench, accuracy increased to 57.5%, showing robust video spatial understanding.
- Ablation studies confirmed that geometric gating and fusion strategies contributed 8-10 points each, while UC-SSRL further enhanced stability and cross-view consistency, validating the importance of geometric consistency in long video reasoning.
Significance
This work addresses fundamental limitations in current multimodal models, which struggle with redundant evidence and view-dependent inconsistencies. By embedding geometric consistency into evidence management and model training, ConsiSpace offers a scalable solution for stable, efficient, and accurate spatial reasoning. Its implications span navigation, robotics, and long video analysis, setting a new standard for multi-view spatial understanding and paving the way for more reliable autonomous systems and intelligent perception modules.
Technical Contribution
The core innovations include the design of the Geometry-Consistent Memory (GCM), which organizes evidence via geometry-guided writing and fusion, and the integration of UC-SSRL, which uses multi-view rewards to optimize cross-view stability. These mechanisms enable the model to filter redundant information, maintain spatial coherence, and adaptively refine reasoning across viewpoints. Unlike prior methods that focus solely on feature fusion or supervised training, ConsiSpace emphasizes evidence lifecycle management grounded in geometric principles, offering theoretical guarantees for stability and efficiency.
Novelty
This is the first work to systematically incorporate geometric consistency as an evidence organization and optimization criterion in long video spatial reasoning. The combination of explicit geometric cues, geometry-gated evidence management, and self-supervised multi-view rewards represents a novel paradigm. Unlike existing geometry-aware models that primarily focus on feature fusion or 3D reconstruction, ConsiSpace emphasizes evidence lifecycle control driven by geometric invariance, enabling more stable and scalable reasoning in dynamic scenes.
Limitations
- While effective, the model's robustness diminishes under extreme viewpoint changes or highly dynamic scenes where geometric cues may be unreliable or inconsistent.
- Dependence on pre-trained geometric encoders like VGGT increases computational costs and may limit performance in complex environments.
- Handling sparse or highly ambiguous evidence remains challenging, indicating room for integrating richer multi-modal cues or adaptive reasoning strategies.
Future Work
Future directions include enhancing geometric encoder robustness, reducing computational overhead, and extending the framework to dynamic scenes with moving objects. Incorporating additional modalities such as LiDAR or radar could further improve spatial understanding. Developing real-time inference capabilities and deploying in real-world navigation and robotics tasks will be critical for translating this research into practical applications.
AI Executive Summary
Video spatial reasoning is vital for autonomous navigation, robotics, and long video comprehension, yet current models struggle with redundant evidence and view-dependent instability. Existing multimodal large language models (MLLMs) tend to prioritize semantic understanding, often neglecting explicit geometric relationships crucial for stable spatial inference. This gap hampers their performance in tasks requiring precise spatial reasoning across long temporal horizons and changing viewpoints.
ConsiSpace introduces a novel framework that embeds geometric consistency into evidence management and model training. Central to this approach is the Geometry-Consistent Memory (GCM), which stores visual and geometric cues—such as pose and depth—using geometry-guided writing, fusion, and hierarchical retrieval strategies. These mechanisms enable the model to efficiently organize spatial evidence, reducing redundancy and enhancing robustness.
A key innovation is the integration of a unified self-supervised reinforcement learning (UC-SSRL) process, which employs multi-view rewards based on answer, metric, and topological consistency. This process refines the model’s cross-view stability without requiring additional annotations, significantly improving answer reliability across different viewpoints.
Extensive experiments on three benchmarks—VSI-Bench, OSI-Bench, and MMSI-Video-Bench—demonstrate that ConsiSpace outperforms existing methods, achieving an average score increase of 12.6 points. Notably, it excels in spatial relation, measurement, and path planning tasks, confirming the effectiveness of geometric consistency in long video reasoning.
This research offers a transformative step toward more stable, efficient, and accurate video spatial reasoning systems. By embedding geometric invariance into evidence organization and leveraging self-supervised multi-view optimization, ConsiSpace paves the way for advanced applications in autonomous navigation, robotic perception, and intelligent video analysis. Future work will focus on extending robustness to dynamic scenes, reducing computational costs, and integrating richer multi-modal cues for broader real-world deployment.
Deep Dive
Glossary
Geometric Consistency (几何一致性)
确保空间关系在不同视角下保持不变的原则（在论文中用于组织和筛选视频证据）; It enforces invariant spatial relations across viewpoints, crucial for stable reasoning.
在模型中作为组织空间证据和优化推理稳定性的核心准则。
Memory-Consistent Reinforcement Learning (记忆一致强化学习)
一种通过多视角奖励机制，优化模型跨视角空间推理稳定性的学习方法; It uses self-supervised rewards to improve multi-view answer, metric, and topological consistency.
在UC-SSRL中实现模型跨视角推理的稳定性。
VGGT (Visual Geometry and Geometric Transformer)
一种编码空间位置、深度和姿态信息的几何编码器，用于提取空间线索; It provides explicit geometric cues like pose and depth for evidence组织。
作为特征提取和证据管理的重要工具。
Gated Writing (几何门控写入)
根据视角变化判断是否写入空间证据，减少冗余; It uses geometric change thresholds to control evidence存储。
在证据管理中实现高效组织。
Open Questions Unanswered questions from this research
- 1 如何在极端动态场景中保持几何一致性仍未充分解决，模型在复杂运动和变形环境下的鲁棒性有待提升。
- 2 多模态几何信息的融合策略尚不完善，如何结合视觉、深度和姿态信息实现更全面的空间理解仍是挑战。
- 3 大规模实际应用中，模型的计算成本和实时性问题未充分解决，需优化架构以适应边缘设备。
Applications
Immediate Applications
自主导航
利用ConsiSpace实现机器人在复杂环境中的稳定路径规划和空间理解，提升自主导航的准确性和鲁棒性。
长视频问答
支持长视频中的空间关系推理，应用于智能助理、视频分析等场景，增强系统的空间认知能力。
Long-term Vision
智能感知系统
结合多模态空间推理，推动自动驾驶、机器人等领域的智能感知系统向真正的空间理解迈进，实现更自主、更智能的环境认知。
Abstract
Video spatial reasoning is essential for navigation-oriented perception and long-video question answering, where models must infer spatial relations across long horizons under changing viewpoints. However, existing multimodal large language models (MLLMs) remain largely semantic-centric, and often fail to reliably aggregate consistent spatial evidence from redundant video observations, leading to inefficient or unstable reasoning. To address these issues, we propose ConsiSpace, a geometry-consistency-aware framework for geometry-sensitive video spatial reasoning that turns spatial consistency into both an evidence organization principle and an explicit post-SFT learning signal. We build a geometry-consistent memory (GCM) including implicit evidence tokens and explicit geometric cues, and leverage efficient organization strategies to compactly preserve task-related spatial evidence. Furthermore, we utilize unified consistency self-supervised reinforcement learning (UC-SSRL) after supervised fine-tuning to improve cross-view stability, with answer-, metric-, and topology-consistency rewards. Extensive experiments on three spatial-reasoning benchmarks, VSI-Bench, OSI-Bench, and MMSI-Video-Bench, show consistent gains, improving the average score by 12.6 points over the strongest baselines.