CoRA Framework for Gradient-Free On-Device Task-Conditioned Retrieval

CoRA enables task-specific in-context learning on edge devices by transforming frozen encoders into retrievers via closed-form ridge regression. It uses low-rank factorization to minimize query-time computation, requiring only the input and a precomputed index without needing gradient updates during inference.