ASIDE: Adaptive and Separable Interventional Dynamics via Progressive Meta-Learning

ICLR 2026 Conference Submission21652 Authors

Keywords: forecasting, dynamics, intervention modelling

Abstract: To inform decisions about changing the future trajectories of a dynamics system, it is important to predict not only the intrinsic dynamics of the system but also its response to external interventions. While notable progress has been made in learning intervention effects over time, existing research has prioritized the challenge of time-varying confounding in observational data. Significant challenges however remain in aspects related to the modeling and inference of latent dynamics. A first and foremost challenge lies in the need to separate, from a composite observation, the natural temporal evolution of intrinsic dynamics from its response to external interventions. This challenge is further exacerbated by the need to integrate rich history information into these latent dynamics. In this paper, we present a novel framework of adaptive and separable interventional dynamics (ASIDE) to overcome these challenges. First, we leverage inductive bias and progressive learning to allow separable modeling and inference of the intrinsic dynamics and its responses to external interventions at the latent space. This is in contrast to existing approaches that model and infer the composite dynamics as a black box. Second, we leverage meta-learning to enable these latent dynamics to adapt to context examples in past history, addressing both inter- and intra-subject variabilities. This is in contrast to existing approaches that use history only to initialize a $\textit{one-size-fit-all}$ forecasting function. On synthetic and real benchmarks, we demonstrate the advantage of ASIDE in improving forecasting accuracy for both intrinsic and interventional dynamics, in settings with or without time-varying confounding.

Primary Area: learning on time series and dynamical systems

Submission Number: 21652