Research map / Vision-language and multimodal / Multimodal LLMs
Unfold The World: Factorize 4D Properties in Reinforcing Spatial Understanding
Yijun Yang, Shenghe Zheng, Wenbo Li, Jianhui Liu, Haoze Sun, Yanbing Zhang, Jiaxiu Jiang, Lin Song, Haoyang Huang, Nan Duan, Lei Zhu
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 T (temporal cycle consistency) reward (ablation) | FactoSR-SFT + GRPO · Camera motion group (Delta_t) | accuracy: 45.7 -> 53.6 (+7.9) | Improved |
| add XY (reprojection correspondence) reward (ablation) | FactoSR-SFT + GRPO · Correspondence group (Delta_Cor) | accuracy: +2.7 (72.6 vs 69.9 SFT) | Improved |
| add Z (depth order) reward (ablation) | FactoSR-SFT + GRPO · Depth group (Delta_Depth) | accuracy: 86.3 -> 87.6 (+1.3) | Improved |
| RL with vanilla GRPO accuracy reward only (ablation) | FactoSR-SFT · 3D/4D average (9 spatial benchmarks) | average accuracy: 60.3 -> 60.5 (+0.2) | No clear effect |
| combine XY, Z and T rewards (ablation) | FactoSR-SFT + GRPO · 3D/4D average (9 spatial benchmarks) | average accuracy: 60.3 -> 62.0 (+1.7) | Improved |
| add vanilla grounding (IoU) reward (ablation) | FactoSR-SFT + GRPO · 3D/4D average (9 spatial benchmarks) | average accuracy: -0.2 overall, -0.7 Spatial_Relation | Hurt |
| use Z reward instead of vanilla grounding reward (ablation) | FactoSR-SFT + GRPO · BLINK Relative_Depth | accuracy: +3.2 (vs +0.8 for grounding) | Improved |
| add Z reward (ablation) | FactoSR-SFT + GRPO · Correspondence group (Delta_Cor) | accuracy: -0.1 (69.8 vs 69.9) | No clear effect |
| combine all rewards vs T reward alone on temporal group (ablation) | FactoSR-SFT + GRPO · Camera motion group (Delta_t) | accuracy: 53.6 -> 50.9 | Mixed |
| add factorized RL stage after SFT | FactoSR-8B · VSI-Bench | average accuracy: 58.5 -> 61.5 | Improved |
| add factorized RL stage after SFT | FactoSR-8B · All-Angles-Bench | average accuracy: 51.6 -> 55.4 | Improved |
| FactoSR training vs base model | Qwen3-VL-8B-Instruct · OCRBench | accuracy: -2.7 | Hurt |
Closest papers by meaning
- Pursuing Minimal Sufficiency in Spatial Reasoning (ICLR 2026 · similarity 0.91)
- ConsisVLA-4D: Advancing Spatiotemporal Consistency in Efficient 3D-Perception and 4D-Reasoning for Robotic Manipulation (CVPR 2026 · similarity 0.90)
- Learning to Reason in 4D: Dynamic Spatial Understanding for Vision Language Models (CVPR 2026 · similarity 0.90)
- From Flatland to Space: Teaching Vision-Language Models to Perceive and Reason in 3D (NeurIPS 2025 · similarity 0.90)
- MindJourney: Test-Time Scaling with World Models for Spatial Reasoning (NeurIPS 2025 · similarity 0.90)
- Think with 3D: Geometric Imagination Grounded Spatial Reasoning from Limited Views (CVPR 2026 · similarity 0.90)
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.
