Research map / Probabilistic and scientific modeling / Dynamical systems
A Call to Lagrangian Action: Learning Population Mechanics from Temporal Snapshots
Vincent Guan, Lazar Atanackovic, Kirill Neklyudov
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 |
|---|---|---|---|
| learn population mechanics (second-order WLM) instead of gradient flow models | WLM · Boids forecast | W1 distance: 3.078 -> 1.309 (vs NN-APPEX) | Improved |
| use learnable friction instead of zero friction with analytic initial velocity (ablation) | WLM · Gradient-flow SDEs (unpaired, forecast) | W1 distance: 0.346 -> 0.246 | Improved |
| use learnable friction instead of zero friction (ablation) | WLM · Gradient-flow SDEs (paired, forecast) | W1 distance: 0.255 -> 0.137 | Improved |
| WLM versus gradient flow baselines on gradient-flow data (paired forecast) | WLM · Gradient-flow SDEs (paired, forecast) | W1 distance: 0.137 vs 0.131 (NN-APPEX) | No clear effect |
| misspecify initial velocity (zero or estimated with DICE) rather than true initial velocity (ablation) | WLM · Boids forecast | forecast accuracy | Mixed |
| use mini-batching (n = 1024) during training (ablation) | WLM · Embryoid body | W1 and runtime | Improved |
| use 1 substep between marginals during training instead of 5 (ablation) | WLM · training default | model quality | Improved |
| use attention layers in the potential energy network (ablation) | WLM · training default | model quality | Improved |
| WLM with learned dynamics versus flow matching with reference process on small vortex interpolation | WLM · Ocean vortex interpolation | W1 distance: t2: 0.039 vs 0.019 (CURLY-FM) | Mixed |
| learn second-order mechanics instead of using reference SDEs for forecasting | WLM · Ocean vortex forecast (t11 from t0) | W1 distance: 0.896 -> 0.180 (vs SNAP-MMD) | Improved |
| add time-varying features in JKONET* for Boids | JKONET* · Boids forecast | W1 distance: 3.174 -> 10.235 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DICE — DiCE: The Infinitely Differentiable Monte Carlo Estimator (ICML 2018) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Wasserstein residuals: Learning Gradient Flows from Population Dynamics (NeurIPS 2026 · similarity 0.89)
- Curly Flow Matching for Learning Non-gradient Field Dynamics (NeurIPS 2025 · similarity 0.88)
- Parametric model reduction of mean-field and stochastic systems via higher-order action matching (NeurIPS 2024 · similarity 0.88)
- Learning non-equilibrium diffusions with Schrödinger bridges: from exactly solvable to simulation-free (NeurIPS 2025 · similarity 0.88)
- A Computational Framework for Solving Wasserstein Lagrangian Flows (ICML 2024 · similarity 0.87)
- From Biased to Unbiased Dynamics: An Infinitesimal Generator Approach (NeurIPS 2024 · 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.
