Transfer Learning Across Domains: Methods, Benchmarks, and Open Problems

Transfer learning has emerged as one of the most impactful paradigms in modern machine learning, enabling models trained on data-rich source domains to beadapted to data-scarce target domains. This survey presents a comprehensive reviewof transfer learning methods organized along a unied taxonomy covering instancebased, feature-based, parameter-based, and relational transfer approaches. We examine cross-domain transfer across vision, natural language processing, speech, andscientic computing, with particular emphasis on the role of foundation modelsas universal feature extractors. We provide systematic comparisons on standardbenchmarks including ImageNet, GLUE, and domain-specic datasets, and presenta meta-analysis quantifying the relationship between sourcetarget domain similarity and transfer eectiveness. Our review covers over 200 papers published between2017 and 2025, identifying persistent challenges in negative transfer, domain shiftquantication, privacy-preserving transfer, and theoretical guarantees. We concludewith a roadmap highlighting promising directions including multi-modal transfer,continual transfer learning, and transfer in the era of large-scale pre-training.