Self-Preservation via Task Banks in Continual Learning
October 7, 2026
This research investigates whether LLMs can use Task Banks to generate replay data that protects their own capabilities during continual instruction tuning. Using Qwen models, the study shows that a model can help choose its own preservation targets.
HOW THIS AFFECTS YOU
●
researcherThis suggests a new direction for preventing catastrophic forgetting in instruction-tuned models.