Title: Have you tried turning it off and on again?
Accelerating molecular dynamics and machine learning with
stochastic resetting
Abstract: In this talk, I will present stochastic resetting as a framework for deciding when interrupting a computation and starting anew can accelerate it. For molecular dynamics, restarting trajectories accelerates rare events without prior knowledge of a reaction coordinate and allows us to infer unbiased mean transition times from accelerated simulations. Combining resetting with existing enhanced sampling method, such as Metadynamics, yields greater acceleration than either method alone. I will also show how machine learning can guide adaptive resetting strategies that respond to the state of the system. For neural-network training, we treat reaching a target accuracy as a first-passage process and use a recently developed response theory [5] to select resetting-based perturbations and their frequency, accelerating the training of classifiers. Finally, I will present Reset-and-Discard for large language models, which reallocates attempts from repeatedly unsuccessful questions to new ones. This strategy increases the number of distinct questions answered within a fixed computational budget and reduces the attempts, tokens, and cost needed to reach a target coverage. Together, these examples connect enhanced sampling of molecular simulations and efficient machine-learning through a common question: when is it better to start over?