Crystal structure prediction (CSP) is essential for discovering materials under extreme conditions, but its application is limited by the prohibitive cost of density functional theory (DFT) relaxations for thousands of candidate structures. Universal machine-learning force fields (MLFFs) provide a much faster alternative, yet their zero-shot accuracy can degrade under high pressure, leading to unreliable candidate ranking. Here we present an entropy-based active-learning strategy that adapts a foundation interatomic potential to the target domain using only a limited set of DFT calculations. We demonstrate the approach for SiO2 by fine-tuning a foundation model on a CSP task at 10 GPa. The resulting model retains first-principles accuracy across a wide range of pressures up to 300 GPa, enabling reliable ranking of candidate structures where the pretrained MLFF fails. Our approach accelerates the CSP pipeline by more than two orders of magnitude while preserving first-principles-level reliability, paving the way for the application of CSP in larger and more complex systems.