Reliable biomarker labels are often scarce in computational pathology, making label-efficient whole-slide image prediction difficult. LFFA introduces a training-time feature-space augmentation framework for frozen slide-level foundation model pipelines, learning controllable augmented slide views that preserve semantic consistency while improving representation diversity. Across three few-shot biomarker prediction tasks and multiple foundation-model backbones, it improves robustness over existing augmentation methods.