Research map / Video / Video-language understanding
ConsiSpace: Learning Geometric Consistency Matters for Video Spatial Reasoning
Ting Huang, zhenyu zhang, Wenyuan Huang, Hao Tang, Jian Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add UC-SSRL consistency RL after SFT | ConsiSpace SFT · OSI-Bench | average: 45.8 -> 53.0 | Improved |
| add UC-SSRL consistency RL after SFT | ConsiSpace SFT · VSI-Bench | average: 71.2 -> 76.6 | Improved |
| add UC-SSRL consistency RL after SFT | ConsiSpace SFT · MMSI-Video-Bench (Uniform-50) | average accuracy: 48.9 -> 58.1 | Improved |
| add geometry-gated writing (ablation) | ConsiSpace SFT · MMSI-Video-Bench (avg of both settings) | accuracy: 47.2 -> 52.3 | Improved |
| add geometry-consistent fusion (ablation) | ConsiSpace SFT · MMSI-Video-Bench (avg of both settings) | accuracy: 47.9 -> 52.3 | Improved |
| add geometry-filtered retrieval (ablation) | ConsiSpace SFT · MMSI-Video-Bench (avg of both settings) | accuracy: 46.0 -> 52.3 | Improved |
| add geometry-filtered retrieval (ablation) | ConsiSpace SFT · VSI-Bench | average: 69.5 -> 71.2 | Improved |
| add all three consistency rewards instead of answer reward only (ablation) | ConsiSpace UC-SSRL · VSI-Bench | average: 73.5 -> 76.6 | Improved |
| add consistency rewards (any single) over plain SFT (ablation) | ConsiSpace SFT · VSI-Bench | average: 71.2 -> 72.8 | Improved |
| add Top-K retrieval to summary-only readout (ablation) | ConsiSpace · VSI-Bench | average: 69.8 -> 71.2 | Improved |
| add summary tokens to Top-K-only readout (ablation) | ConsiSpace · OSI-Bench | average: 46.2 -> 45.8 | Mixed |
| raise writing threshold tau_p from 0.15 to 0.25 m (ablation) | ConsiSpace SFT · VSI-Bench | average: 71.2 -> 70.9 | No clear effect |
Closest papers by meaning
- The Art of Interrogation: Consistency Amplifies Factuality in Spatial Reasoning (ICML 2026 · similarity 0.91)
- ConsisVLA-4D: Advancing Spatiotemporal Consistency in Efficient 3D-Perception and 4D-Reasoning for Robotic Manipulation (CVPR 2026 · similarity 0.89)
- Thinking with Geometry: Active Geometry Integration for Spatial Reasoning (ICML 2026 · similarity 0.89)
- SpatialLadder: Progressive Training for Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.89)
- Video Spatial Reasoning with Object-Centric 3D Rollout (AAAI 2026 · similarity 0.89)
- Spatial-DISE: A Unified Benchmark for Evaluating Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
