Generative Active Testing Improves Benchmark Sample Selection

Researchers led by Aashish Anantha Ramakrishnan present Generative Active Testing (GAT) in an arXiv preprint dated Feb 26, 2026, introducing an uncertainty-aware acquisition framework that uses LLMs as surrogates for sample selection in generative QA benchmarks. GAT's Statement Adaptation Module converts generative tasks into pseudo-classification, enabling zero-shot acquisition functions that cut estimation error by about 40% versus traditional sampling baselines, reducing labeling costs for expert-annotated domains.