Research map / Probabilistic and scientific modeling / Dynamical systems
Which priors matter? Benchmarking models for learning latent dynamics
Aleksandar Botev, Andrew Jaegle, Peter Wirnsberger, Daniel Hennes, Irina Higgins
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 |
|---|---|---|---|
| model dynamics with Neural ODE with time-reversal constraint instead of Hamiltonian prior | HGN · 17 datasets | forward VPT best-in-dataset count: 11/17 | Improved |
| model dynamics with Neural ODE with time-reversal constraint instead of Hamiltonian prior | HGN · 17 datasets | backward VPT best-in-dataset count: 13/17 | Improved |
| model dynamics continuously (Neural ODE instead of discrete RGN Res) (ablation) | RGN Res · 17 datasets | forward VPT average rank: 3.0 -> 2.88 | Improved |
| use Hamiltonian prior instead of Lagrangian prior (ablation) | LGN NEW · 17 datasets | backward VPT average rank: 2.12 -> 2.59 | Hurt |
| use Lagrangian prior instead of Hamiltonian prior (ablation) | HGN · 17 datasets | forward VPT average rank: 4.47 -> 4.18 | Mixed |
| Lagrangian prior needs matrix inversion for acceleration (ablation) | LGN · training cost | per-step compute | Hurt |
| predict residual state updates (RGN Res instead of RGN) (ablation) | RGN · 17 datasets | forward VPT average rank: 5.65 -> 3.0 | Improved |
| residual prediction (RGN Res instead of autoregressive AR) (ablation) | AR · 17 datasets | forward VPT average rank: 5.65 -> 3.0 | Improved |
| add time-reversal constraint to Neural ODE (ODE[TR] vs ODE) (ablation) | ODE · 17 datasets | backward VPT average rank: 3.47 -> 1.41 | Improved |
| add time-reversal constraint to Neural ODE (ODE[TR] vs ODE) (ablation) | ODE · 17 datasets | forward VPT average rank: 2.88 -> 2.47 | Improved |
| add friction and control to mass-spring (+c +f vs plain) (ablation) | HGN · Mass-spring | forward VPT: 1.00(0.00) -> 0.41(0.17) | Hurt |
| increase MD system size from 4 to 16 particles (ablation) | all models · MD | forward VPT: 0.28(0.05) -> 0.00(0.00) | Hurt |
Closest papers by meaning
- Generating Physical Dynamics under Priors (ICLR 2025 · similarity 0.87)
- Identifying latent state transitions in non-linear dynamical systems (ICLR 2025 · similarity 0.86)
- Learning Hybrid Dynamics Models with Simulator-Informed Latent States (AAAI 2024 · similarity 0.86)
- Latent Mixture of Symmetries for Sample-Efficient Dynamic Learning (NeurIPS 2025 · similarity 0.86)
- Learning Physics From Video: Unsupervised Physical Parameter Estimation for Continuous Dynamical Systems (CVPR 2025 · similarity 0.85)
- ODEWorld: A Continuous Predictive Architecture via Physical-Time Flow (NeurIPS 2026 · similarity 0.85)
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.
