Negative Transfer

Negative transfer studies when sharing learned features, parameters, data, or skills makes a target task perform worse than learning it independently. Researchers address it by screening or reweighting sources, clustering related tasks, and routing updates through specialized experts or lightweight adapters; these approaches appear in multi-task learning, federated learning, domain adaptation, and continual learning. Evaluation compares transfer with target-only baselines and increasingly tracks asymmetry, held-out performance, and person- or task-level losses rather than averages alone. The central challenge is deciding when knowledge will help: related tasks can still interfere, and isolation that prevents harm can also discard useful sharing.