Video walkthrough
Follow an agent-guided optimization experiment and see how observations, requested changes, and subsequent measurements connect.
Feedback-driven neural network optimization
Steer a running training job.
Interactive Training connects live metrics to interventions by humans and AI agents. Observe a run, request a change, and inspect what happened when the training process applied it.
Three ways to explore the system
Recorded paper trace
Compare training rounds, configurations, scores, actions, and the reflections carried into the next round. This view reads a recorded experiment.
Live CPU sandbox
Join the shared CPU queue for a reduced tiny-BERT training run. Explore its exposed controls and their recorded outcomes. This sandbox uses no LLM calls.
From observation to intervention
A controller reads training measurements and submits changes through a control interface. The training process applies those requests while the run continues. Optimizer settings, data handling, and checkpoints become part of an interactive feedback loop.
Interactive Training 2 makes this interface a shared protocol. A training application declares typed settings and actions, applies requests at its control points, and records requests and outcomes in an ordered journal. Humans, scripts, and language-model agents use the same interface.
Papers and authors
Interactive Training: Feedback-Driven Neural Network Optimization
The original framework connects live monitoring to feedback-driven intervention in neural-network training, with case studies on stability, sensitivity to initial hyperparameters, and changing user needs.
Interactive Training 2: Auditable Control Plane for Live Model Training
A reusable control protocol and an Aim workspace connect live metrics, settings, actions, and their execution record across five NLP and reinforcement-learning workflows.
Methods and connections
Controlling a running training job with humans and AI agents
Read the research guide for the method, its relationship to live hyperparameter tuning and training agents, and a bibliography with reusable citations.
Read the research guide →