Learning, Fast and Slow: Towards LLMs That Adapt Continually
The paper introduces Fast-Slow Training (FST) to adapt Large Language Models continually. It separates task-specific information into 'fast' weights (optimized context/prompts) and 'slow' weights (persistent model parameters). FST is more sample-efficient, preserves plasticity against forgetting, and allows the model to learn new tasks on the fly while maintaining its general reasoning abilities.