Tag
This paper identifies two distinct failures in learned physical simulators—instability in long rollouts and inability to adapt to changed laws—and proposes separate structural solutions: symplectic integration for stability and explicit factorization for counterfactual generalization.
This paper presents a memory–stability–expressivity trilemma for trainable dissipative oscillator networks, showing that damping governs all three and limits trainability, with experimental validation on a 20-oscillator network confirming the theoretical bounds.