Dynamically deciding when to halt refinement based on learned signals rather than using a fixed number of steps.