Generating training data using the current model being trained, rather than a fixed dataset, to ensure relevance to the model's current behavior.