ConsiSpace: Learning Geometric Consistency Matters for Video Spatial Reasoning

2026-07-20

AI Summary

Current vision-language models struggle to effectively aggregate cross-view consistent spatial evidence in long-form multi-view video reasoning, leading to inefficiency and instability. This work proposes ConsiSpace, a novel framework that uniquely leverages geometric consistency both as a principle for organizing spatial evidence and as an explicit learning signal. The approach introduces a Geometrically Consistent Memory (GCM) architecture, an efficient strategy for spatial evidence organization, and a Unified Consistency Self-Supervised Reinforcement Learning (UC-SSRL) mechanism. Reasoning is further refined through a tripartite reward system based on answer correctness, metric consistency, and topological coherence. Evaluated on VSI-Bench, OSI-Bench, and MMSI-Video-Bench, ConsiSpace achieves an average improvement of 12.6 points over the strongest existing baselines, demonstrating significant gains in performance and robustness.

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.

Problem

• video spatial reasoning
• geometric consistency
• multimodal large language models
• spatial evidence aggregation
• cross-view stability

Innovation

• geometric consistency
• video spatial reasoning
• geometry-consistent memory
• self-supervised reinforcement learning
• multimodal large language models