Abstract
Generative artificial intelligence is transforming materials discovery by creating unexplored candidates. However, such models are trapped in historical data bias, particularly for data-scarce systems like two-dimensional (2D) materials, leading to repetitive outputs rather than genuine discoveries. Here, we introduce DuALGen, a dual active learning framework that mitigates these limitations. DuALGen couples two complementary loops to enrich data diversity and correct data bias: a generative loop that uses dynamic, multi-criteria sampling to drive exploration of the design space, and a predictive loop that samples outliers to counter distribution shift, enabling reliable evaluation of novel, previously unknown candidates. Applied to 2D materials, DuALGen uncovers >10 000 stable, distinct compounds, including thousands of high-performance candidates for electronic applications. This self-updating workflow connects generative models to uncharted chemical spaces, and offers a practical route to continuous discovery of new materials.