Reactive Optimizers

We put an LLM agent at the heart of the optimizer. It watches the training dynamics, reacts to them, and rewrites the optimizer’s source code as training goes, all within a single run. When no recipe has been tuned for a task, our system OPAL (OPtimizer with an Agent in the Loop) matches or beats the best off-the-shelf recipe across image and language tasks. The agent spontaneously discovers sophisticated schedules, such as learning-rate increases and per-module rate changes, and intricate strategies to prevent loss spikes and escape overfitting.