Query synthesis is an active learning approach where, instead of picking existing unlabeled data, the model generates synthetic instances to be labeled​. The idea is to synthesize queries that would be maximally informative for the model. For example, the model might create an input (or perturb an existing one) that it finds particularly confusing and ask an oracle for its label. These methods leverage the model’s current state to propose new data points that could most reduce its uncertainty or error​.Techniques in query synthesis include generating points near decision boundaries or using models to hallucinate examples. By incorporating synthetic yet informative examples, query synthesis can improve model performance while still minimizing the need for real labeled data​.