New Framework Improves Transfer in Continual Learning Models

This research challenges the conventional focus of continual learning, which primarily aims to prevent catastrophic forgetting in language models. Instead, the authors argue for treating "forward transfer"—where past knowledge aids future learning—as a primary objective. They introduce a framework that defines three conditions necessary for effective transfer: room for improvement in the target task, survival of transferable information during optimization, and compatibility of replay data from previous tasks. To instantiate this framework, the paper presents Transfer-Selective Replay (TSR). Unlike indiscriminate replay methods, TSR intelligently selects past data predicted to be most beneficial for the incoming task, guided by a zero-training task signature. This selective approach, combined with distillation to preserve stability on prior tasks, consistently enhances forward transfer while maintaining stability, particularly in low-budget continual learning scenarios. The findings, demonstrated across heterogeneous and homogeneous task streams, advocate for a more deliberate and structured approach to understanding and engineering transfer mechanisms in continual learning, moving beyond solely preventing forgetting.