Interactive Training 2: Auditable Control Plane for Live Model Training
Wentao Zhang, Xuanhe Pan, Han Zhou, Yang Lu et al. · arXiv · 2026
Abstract
Experiment trackers show what a training run is doing but not how to change it — adjusting a live run still means trainer-specific code. This is an open-source control plane where the training application declares which settings and actions it exposes, humans and automated controllers submit requests through one shared interface, and the loop applies them at safe points. Every request and its outcome is recorded chronologically beside the metrics.
Why it matters
The auditability is the point. Once agents can steer training runs, a written record of who changed what and what happened next stops being a nicety and becomes the only way to debug.
https://arxiv.org/abs/2607.18314