Tag
A fast-slow learning framework for LLMs combines fixed slow weights with optimized fast context weights, achieving up to 3x better sample efficiency and reduced catastrophic forgetting in continual learning scenarios.