AI Learns to Sleep: Google Wants to Cure Large Models' Chronic Forgetting-While-Learning Problem

Large models are fixed as soon as they leave the factory. New knowledge is either lost the moment a conversation ends, or it is crammed in and causes the model to forget old capabilities. On October 27, Google released the "Sleep Framework," and the paper "Language Models Need Sleep" proposed that models should consolidate memories through sleep like humans, solving the "knowledge cutoff date" problem. The framework lets models work while awake and organize while asleep, balancing learning new things and preserving old ones by switching operating states. It was released by Google Research.