Interactive Training 2 opens auditable control plane for ML runs

There is a persistent annoyance in ML infrastructure: you can watch a training run progress in real time on any experiment tracker, but if you actually want to change something mid-flight, you are back to trainer-specific patches. A new paper on arXiv from Wentao Zhang, Xuanhe Pan, Han Zhou, Yang Lu, and Yuntian Deng proposes a shared protocol as the fix, called Interactive Training 2. Training applications declare which settings and actions they expose, humans and automated controllers submit requests through the same interface, and the training loop validates and applies them at safe control points.

Two things about the framing stand out. First, humans and agents are first-class citizens of the same API. That matters because agent-guided training, a controller that watches loss curves and tunes settings on its own, has been sitting in the space between research demo and production because there was no clean way to wire it in without forking the trainer. Second, every request and its outcome lands in a chronological record inside a customized Aim workspace, which is the piece that turns "we hot-patched the LR" into an auditable event rather than a support-ticket reconstruction.

The authors demonstrate the system across five NLP and reinforcement-learning workflows: a BERT sentiment classifier, a multi-source sentiment mixing setup, a layerwise GPT that exposes 27 embedding, block, head, and residual-group learning rates as independent controls, a Muon-AdamW GPT, and an RLVR Countdown reinforcement-learning task. They report the RLVR task's hardest-split accuracy going from 0.032 to 0.232, and the Muon-AdamW GPT validation loss from 4.797 to 4.429. Take those as demonstrations of what the interface makes possible, not as evidence that live tuning generally beats a good static schedule.

The honest caveat is that this is a system release rather than a controlled study, so the paper does not tell you when live intervention actually pays off versus a well-tuned baseline, nor how the safe-control-point abstraction behaves under distributed frameworks like FSDP or DeepSpeed. What is interesting is the release itself, published under github.com/yuntian-group/interactive-training and a live demo at interactivetraining.ai/live: a shared substrate for anyone wanting to build closed-loop agent trainers, or run models in an environment where "who changed what during this training run" needs to be a real, queryable answer.