| title | ConsiSpace: Learning Geometric Consistency Matters for Video Spatial Reasoning |
|---|---|
| authors | Ting Huang, Zhenyu Zhang, Wenyuan Huang, Jian Yang, Hao Tang |
| date | 2026-07-20 |
| venue | arXiv (preprint, not peer reviewed) |
| id | arXiv:2607.17599 |
| upvotes | 2 (Hugging Face Daily Papers, 2026-11-01) |
| links | abs pdf hf |
| 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. |
cite
@misc{huang2026consispace,
title={ConsiSpace: Learning Geometric Consistency Matters for Video Spatial Reasoning},
author={Ting Huang and Zhenyu Zhang and Wenyuan Huang and Jian Yang and Hao Tang},
year={2026},
eprint={2607.17599},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2607.17599}
}