OPAL lets agents rewrite optimizers mid-training

OPAL makes optimization feel less like selecting a static recipe and more like supervising an adaptive programmer—but running mutable code inside training introduces serious reproducibility and safety tradeoffs.

• –Agents can alter learning rates, momentum, preconditioning, clipping, and parameter-group behavior as conditions change
• –The approach responds to loss spikes, corrupted data, and overfitting without optimizer-specific meta-training
• –It preserves the familiar PyTorch workflow through a lightweight optimizer wrapper
• –Agent calls can add substantial runtime and inference cost, especially for short training jobs
• –Because rewritten programs execute inside the training process, failures can degrade a run even with limited rollback support