TL;DR of the paper, generated automatically.

Paper proposes a training method for learning dynamical systems without simulating them during training.

• Method: Velocity Matching, a simulation-free objective for learning continuous-time dynamics.
• Claim: avoids long-rollout instability and gives better accuracy on ODEs and PDEs, per the abstract.
• Caveat: the numbers come from the paper's own experiments on nonlinear ODEs and PDEs.