A Minimal Interpretable Architecture for Zero-Shot Reconstruction of Dynamical Systems

arXiv cs.AI · 6 hours ago

Researchers reduced a state-of-the-art foundation model for dynamical systems reconstruction called DynaMix to a minimal two-parameter model named DynaBase that uses linear blending of current latent state and nearest in-context neighbors. DynaBase achieves competitive zero-shot forecasting performance across chaotic and cyclic systems with parameter counts many orders of magnitude lower than other foundation models. The extreme simplicity of DynaBase enables direct model optimization and closed-form analytical solutions, revealing the minimal mechanisms required for effective zero-shot dynamical systems reconstruction.