Research map / Vision-language and multimodal / Multimodal LLMs
VLM4D: Towards Spatiotemporal Awareness in Vision Language Models
Shijie Zhou, Alexander Vilesov, Xuehai He, Ziyu Wan, Shuwang Zhang, Aditya Nagachandra, Di Chang, Dongdong Chen, Xin Eric Wang, Achuta Kadambi
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 |
|---|---|---|---|
| supervised fine-tuning on real VLM4D training split (ablation) | Qwen2-VL-7B · VLM4D real test split | accuracy: 38.3 -> 54.5 | Improved |
| supervised fine-tuning on real VLM4D training split (ablation) | Qwen2.5-VL-7B · VLM4D real test split | accuracy: 43.4 -> 55.6 | Improved |
| add synthetic data to real data for SFT (ablation) | Qwen2.5-VL-7B · VLM4D real test split | accuracy: 55.6 -> 56.3 | No clear effect |
| add synthetic data to real data for SFT (ablation) | Qwen2-VL-7B · VLM4D real test split | accuracy: 54.5 -> 55.5 | No clear effect |
| fine-tune on synthetic data only (ablation) | Qwen2-VL-7B · VLM4D real test split | accuracy: 38.3 -> 42.5 | Improved |
| use reconstructed 4D global feature field as input | InternVideo2-8B · VLM4D DAVIS 2016 (Direct Output) | accuracy: 24.3 -> 29.0 | Improved |
| use reconstructed 4D global feature field as input | InternVideo2-8B · VLM4D DAVIS 2016 (CoT) | accuracy: 36.0 -> 37.4 | Improved |
| use rendered global-view video instead of original 2D video (ablation) | InternVideo2-8B · VLM4D DAVIS 2016 (CoT) | accuracy: 36.0 -> 32.7 | Hurt |
| chain-of-thought prompting instead of direct output (ablation) | evaluated VLMs · VLM4D | accuracy | No clear effect |
| dense captions in video SFT datasets contain spatiotemporal terms | ShareGPT4Video · manual check of 100 labels | label accuracy: less than 10% | Hurt |
Closest papers by meaning
- Linear Mechanisms for Spatiotemporal Reasoning in Vision Language Models (ICLR 2026 · similarity 0.92)
- MLLM-4D: Towards Visual-based Spatial-Temporal Intelligence (ICML 2026 · similarity 0.92)
- ConsisVLA-4D: Advancing Spatiotemporal Consistency in Efficient 3D-Perception and 4D-Reasoning for Robotic Manipulation (CVPR 2026 · similarity 0.92)
- Thinking in Dynamics: How Multimodal Large Language Models Perceive, Track, and Reason Dynamics in Physical 4D World (CVPR 2026 · similarity 0.91)
- GTR-Bench: Evaluating Geo-Temporal Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.91)
- OST-Bench: Evaluating the Capabilities of MLLMs in Online Spatio-temporal Scene Understanding (NeurIPS 2025 · similarity 0.91)
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.
