ConsiSpace | Video Spatial Reasoning

ConsiSpace · Research presentation04:59 · Sound on

A walkthrough of the motivation, memory design, consistency learning, and experiments. The sections below provide a text overview.

Overview

Consistent evidence, from memory to learning.

Existing multimodal models often retain redundant video observations and produce unstable answers across viewpoints. ConsiSpace uses 3D spatial consistency to compact evidence and turn cross-view agreement into an explicit learning signal.

ConsiSpace overview: compact memory, cross-view stability, and benchmark performance. — Figure 1. Efficiency, robustness, and performance of ConsiSpace. View full resolution ↗

Method

Two ideas, one consistency principle.

ConsiSpace architecture with visual and spatial encoders, dual memory, query fusion, and consistency rewards. — Figure 2. Geometry-Consistent Memory and unified consistency learning in one framework. View full resolution ↗

01 · MEMORY

Geometry-Consistent Memory

Implicit visual-spatial evidence and explicit pose-depth cues are managed through geometry-gated writing, consistent fusion, and hierarchical retrieval.

02 · LEARNING

Unified Consistency SSRL

Answer, metric, and topology rewards refine cross-view stability after supervised fine-tuning without requiring additional human annotations.

Results

Strong gains across three benchmarks.

An average gain of 12.6 points across the four evaluation settings below.

Overall scores reported in Tables 1–3 of the paper. Higher is better.
Benchmark	Setting	Strongest baseline	ConsiSpace	Gain
VSI-Bench	Overall	69.6	76.6	+7.0
OSI-Bench	Overall	40.3	53.0	+12.7
MMSI-Video-Bench	Sufficient-Coverage	42.6	57.5	+14.9
MMSI-Video-Bench	Uniform-50	43.1	58.1	+15.0

Scores use each benchmark’s evaluation protocol. See full experimental results ↗

Qualitative results

Different views. Consistent answers.

Revisited views of trash bins: baseline answers vary, while ConsiSpace answers four in both views. — Figure 7. Answer consistency when revisiting the same scene. View full resolution ↗

If ConsiSpace supports your research, please cite our work.

@article{huang2026consispace,
  title  = {ConsiSpace: Learning Geometric Consistency Matters
            for Video Spatial Reasoning},
  author = {Huang, Ting and Zhang, Zhenyu and Huang, Wenyuan
            and Yang, Jian and Tang, Hao},
  journal = {arXiv preprint arXiv:2607.17599},
  year   = {2026},
  doi    = {10.48550/arXiv.2607.17599},
  url    = {https://arxiv.org/abs/2607.17599}
}