Wentao Zhang Yang Young Lu Yuntian Deng

Abstract

Traditional neural network training typically follows fixed, predefinedoptimization recipes, lacking the flexibility to dynamically respond toinstabilities or emerging training issues. In this paper, we introduceInteractive Training, an open-source framework that enables real-time,feedback-driven intervention during neural network training by human experts orautomated AI agents. At its core, Interactive Training uses a control server tomediate communication between users or agents and the ongoing training process,allowing users to dynamically adjust optimizer hyperparameters, training data,and model checkpoints. Through three case studies, we demonstrate thatInteractive Training achieves superior training stability, reduced sensitivityto initial hyperparameters, and improved adaptability to evolving user needs,paving the way toward a future training paradigm where AI agents autonomouslymonitor training logs, proactively resolve instabilities, and optimize trainingdynamics.