Interactive Training

Interactive Training enables humans and AI agents to monitor neural network training and intervene while a run is in progress. Its control server connects live training metrics to changes in optimizer hyperparameters, training data, and model checkpoints.

Interactive Training 2 provides a shared protocol for human and automated control of live training. Applications register settings and actions, apply requests at explicit control points, and record requests and outcomes.

Open the interactive demo to explore the recorded paper trace and a live CPU training sandbox. Enable JavaScript to use the interactive controls.

Papers and code

Interactive Training: Feedback-Driven Neural Network Optimization. Wentao Zhang, Yang Young Lu, and Yuntian Deng. EMNLP 2025 System Demonstrations.
Interactive Training 2: Auditable Control Plane for Live Model Training. Wentao Zhang, Xuanhe Pan, Han Zhou, Yang Lu, and Yuntian Deng. 2026.
Source code and installation instructions.