Algorithms sampling from thermal states of many-body systems often get stuck in restricted, metastable states. Here, authors prove that the true underlying models can nevertheless be accurately recovered from these trapped samples, establishing rigorous learning guarantees for discrete distributions.