Research map / Video / Video-language understanding
Dynamic Visual Reasoning by Learning Differentiable Physics Models from Video and Language
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| optimize collision-independent parameters first using pre-collision frames (ablation) | VRDP physics fitting · CLEVRER counterfactual | accuracy per question: 49.6 -> 74.1 | Helped |
| add curriculum optimization over increasing frame ranges for mass and restitution (ablation) | VRDP physics fitting · CLEVRER counterfactual | accuracy per question: 74.1 -> 84.3 | Helped |
| add re-optimization on last 20 frames for prediction (ablation) | VRDP physics fitting · CLEVRER predictive | accuracy per question: 76.5 -> 91.4 | Helped |
| add re-optimization on last 20 frames (ablation) | VRDP physics fitting · CLEVRER counterfactual | accuracy per question: 84.3 -> 84.3 | No clear effect |
| use supervised Faster R-CNN detector instead of unsupervised slot attention (ablation) | VRDP · CLEVRER overall | accuracy per question: 86.9 -> 89.4 | Helped |
| add object attribute supervision and rule-based executor (ablation) | VRDP · CLEVRER counterfactual | accuracy per question: 80.7 -> 84.3 | Helped |
| add differentiable physics engine vs end-to-end attention model | Object-based Attention · CLEVRER counterfactual | accuracy per question: 75.6 -> 80.7 | Helped |
| add differentiable physics vs end-to-end attention model, explanatory | Object-based Attention · CLEVRER explanatory | accuracy per question: 96.0 -> 85.2 | Hurt |
| use VRDP vs baselines on few-shot new concepts | Object-based Attention · Generalized CLEVRER | accuracy per question: 26.7 -> 75.6 | Helped |
| use VRDP vs CIN on long-term prediction | CIN · Real-Billiards | S2 error (x1000): 2.34 -> 0.88 | Helped |
Closest papers by meaning
- Dynamics: Language-Based Representation for Inferring Rigid-Body Dynamics From Videos (CVPR 2026 · similarity 0.89)
- Beyond Static Vision: Scene Dynamic Field Unlocks Intuitive Physics Understanding in Multi-modal Large Language Models (ICLR 2026 · similarity 0.89)
- Latent Intuitive Physics: Learning to Transfer Hidden Physics from A 3D Video (ICLR 2024 · similarity 0.88)
- Learning 3D Particle-based Simulators from RGB-D Videos (ICLR 2024 · similarity 0.87)
- VideoREPA: Learning Physics for Video Generation through Relational Alignment with Foundation Models (NeurIPS 2025 · similarity 0.86)
- QUANTIPHY: A Quantitative Benchmark Evaluating Physical Reasoning Abilities of Vision-Language Models (CVPR 2026 · similarity 0.86)
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.