Research map / Probabilistic and scientific modeling / Dynamical systems
On the Learning Mechanisms in Physical Reasoning
Shiqian Li, Kewen Wu, Chi Zhang, Yixin 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 |
|---|---|---|---|
| Use ViT to classify outcomes directly from initial scenes instead of predicting object dynamics with RPIN | PHYRE-B physical reasoning · PHYRE-B cross-template | AUCCESS: 50.86 -> 56.31 | Improved |
| Use ViT direct classification instead of RPIN dynamics prediction | PHYRE-B physical reasoning · PHYRE-B within-template | AUCCESS: 85.49 -> 84.16$\pm$0.30 | Mixed |
| Supply eight ground-truth dynamics frames instead of intuition-only input (ablation) | TimeSformer on PHYRE-B · PHYRE-B within-template and cross-template | AUCCESS: +23.11 within; +48.16 cross | Improved |
| Replace ground-truth simulator dynamics with PredRNN-predicted frames using parallel optimization (ablation) | TimeSformer on PHYRE-B · PHYRE-B within-template and cross-template | AUCCESS: Simulator 4 frames: 86.75, 73.81; PredRNN parallel: 75.22, 46.42 | Hurt |
| Train PredRNN serially and freeze it before training TimeSformer, instead of parallel optimization (ablation) | PredRNN + TimeSformer on PHYRE-B · PHYRE-B within-template and cross-template | AUCCESS: 75.22 -> 64.90 within; 46.42 -> 44.33 cross | Hurt |
| Use predicted future frames rather than a single initial frame with TimeSformer (ablation) | TimeSformer on PHYRE-B · PHYRE-B within-template | AUCCESS: 73.53 -> 75.22 | Improved |
| Use predicted future frames rather than a single initial frame with TimeSformer (ablation) | TimeSformer on PHYRE-B · PHYRE-B cross-template | AUCCESS: 46.02 -> 46.42 | Mixed |
| Use Swin instead of ViT for direct image classification (ablation) | LFI on PHYRE-B · PHYRE-B within-template | AUCCESS: 84.16$\pm$0.30 -> 84.71$\pm$0.33 | Improved |
| Use Swin instead of ViT for direct image classification (ablation) | LFI on PHYRE-B · PHYRE-B cross-template | AUCCESS: 56.31$\pm$1.95 -> 54.92$\pm$2.30 | Hurt |
| Use BEiT instead of ViT for direct image classification (ablation) | LFI on PHYRE-B · PHYRE-B within-template | AUCCESS: 84.16$\pm$0.30 -> 83.59$\pm$0.09 | Hurt |
| Use BEiT instead of ViT for direct image classification (ablation) | LFI on PHYRE-B · PHYRE-B cross-template | AUCCESS: 56.31$\pm$1.95 -> 54.07$\pm$1.88 | Hurt |
Closest papers by meaning
- Causal-PIK: Causality-based Physical Reasoning with a Physics-Informed Kernel (ICML 2025 · similarity 0.85)
- Scaling Physical Reasoning with the PHYSICS Dataset (NeurIPS 2025 · similarity 0.84)
- ContPhy: Continuum Physical Concept Learning and Reasoning from Videos (ICML 2024 · similarity 0.83)
- Demystifying Reasoning Dynamics with Mutual Information: Thinking Tokens are Information Peaks in LLM Reasoning (NeurIPS 2025 · similarity 0.83)
- I-PHYRE: Interactive Physical Reasoning (ICLR 2024 · similarity 0.83)
- Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning (ICML 2026 · similarity 0.83)
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.
