Dynamics-Informed Weight Diffusion for Generalizable Prediction of Complex Systems

ICLR 2026 Conference Submission15164 Authors

Keywords: Dynamical system, Multi-environment learning, Diffusion

TL;DR: We propose a novel generative meta-learning method for generalizable prediction of  complex systems.

Abstract: Data-driven methods offer an effective equation-free solution for predicting physical dynamics. However, predictive models often fail to generalize to unseen environments due to varying dynamic behaviors. In this work, we introduce DynaDiff, a novel generative meta-learning framework to enable efficient, test-time adaptation. Instead of tuning a pre-trained model or context, DynaDiff directly generates a complete, high-performance expert model from scratch, conditioned on a short observation sequence from a new target environment. Specifically, we first finetune a base model on various source environments to efficiently construct a model zoo of expert predictors. Subsequently, we leverage a weight graph representation and train a conditional diffusion model to learn the underlying distribution of expert weights, capable of generating new models from a given dynamic behavior. To effectively capture the dynamic context from the observation sequence, we design a dynamics-informed prompter that explicitly models the relationship between the system's state and its temporal evolution, providing a highly informative prompt for the generative process. Extensive experiments demonstrate that our method can generate expert models with strong generalization for new environments, conditioned on limited observations.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 15164