Research map / Video / Video-language understanding
Reason, Then Re-reason: Cross-view Revisiting Improves Spatial Reasoning
Chaofan Ma, Zhenjie Mao, Yuhuan Yang, Fanqin Zeng, Yue Shi, Yingjie Zhou, Xiaofeng Cao, Jiangchao Yao
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 Re-reason phase with synthesized novel-view video | Qwen3-VL-4B · VSI-Bench | average accuracy: 30.7 -> 36.5 | Improved |
| Add ReRe to Qwen3-VL-2B on STI-Bench static | Qwen3-VL-2B · STI-Bench static subset | average accuracy: 22.2 -> 30.2 | Improved |
| Add ReRe to InternVL2.5-4B | InternVL2.5-4B · STI-Bench static subset | average accuracy: 30.5 -> 30.6 | No clear effect |
| Sequential revisiting instead of concatenating the two videos (ablation) | Qwen2.5-VL-7B · VSI-Bench | average accuracy: 25.4 -> 29.5 | Improved |
| Sequential revisiting instead of interleaved joint reasoning (ablation) | Qwen2.5-VL-7B · VSI-Bench | average accuracy: 25.6 -> 29.5 | Improved |
| Re-reason on the same original video instead of the novel view (ablation) | Qwen2.5-VL-7B · VSI-Bench | average accuracy: 24.8 -> 23.5 | Hurt |
| Use synthesized novel view alone for both phases instead of combining with egocentric video (ablation) | Qwen2.5-VL-7B · VSI-Bench | average accuracy: 29.5 -> 27.3 | Hurt |
| Use Oblique Sweep trajectory instead of Mid-level Traverse (ablation) | ReRe · VSI-Bench | average accuracy: 25.6 -> 29.5 | Improved |
| Use Oblique Sweep trajectory instead of Bird's-eye Orbit (ablation) | ReRe · VSI-Bench | average accuracy: 27.4 -> 29.5 | Improved |
| Reduce VGGT input frames from 100 to 20 (ablation) | ReRe · VSI-Bench | average accuracy: 35.8 -> 33.3 | Mixed |
| Add ReRe to spatial-specialized SpaceR-3B | SpaceR-3B · VSI-Bench | average accuracy: 34.69 -> 35.96 | Improved |
| Add ReRe to spatial-specialized SpatialLadder-3B | SpatialLadder-3B · VSI-Bench | average accuracy: 44.84 -> 45.58 | Improved |
Closest papers by meaning
- ReVSI: Rebuilding Visual Spatial Intelligence Evaluation for Accurate Assessment of VLM 3D Reasoning (ICML 2026 · similarity 0.89)
- 3ViewSense: Spatial and Mental Perspective Reasoning from Orthographic Views in Vision-Language Models (ICML 2026 · similarity 0.88)
- Thinking with Geometry: Active Geometry Integration for Spatial Reasoning (ICML 2026 · similarity 0.87)
- Video Spatial Reasoning with Object-Centric 3D Rollout (AAAI 2026 · similarity 0.87)
- SpaceVista: All-Scale Visual Spatial Reasoning from mm to km (ICML 2026 · similarity 0.87)
- SpinBench: Perspective and Rotation as a Lens on Spatial Reasoning in VLMs (ICLR 2026 · similarity 0.87)
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.
