Minimal yet Mighty adaptive Cosine Projection (MaCP) is a new efficient adaptation method proposed by the University of Amsterdam on May 29, 2025, which aims to achieve excellent performance in fine-tuning large base models with minimal parameters and memory overhead.MaCP: Minimal yet Mighty Adaptation via Hierarchical Cosine Projection", which won the ACL 25 Best Theme Paper Award.

MaCP leverages the energy compression and decorrelation properties of the discrete cosine transform (DCT) to project weight changes in low-rank adaptation into discrete cosine space and select the most critical frequency components at multiple frequency levels, thereby improving model efficiency and accuracy. Compared to existing efficient parameter fine-tuning methods (such as LoRA, VeRA, and LaMDA) on multiple tasks, MaCP significantly reduces memory usage and computational complexity while improving accuracy.