Synthetic Data Selection

The field asks which generated examples are useful for a particular learner and task, across settings such as language-model training, image recognition, and tabular prediction. Methods rank or filter samples using signals such as task utility, realism, diversity, embedding coverage, decision-boundary value, and feedback from model training; some also select groups of examples or guide generation toward underrepresented regions. These approaches can reduce the amount of synthetic data needed, but the best signals and selection rules vary with the generator, training budget, and target task. A central open tension is that convenient proxies—such as provenance, uncertainty, or similarity to real data—do not always predict downstream benefit, so evaluation against the intended training objective remains important.