The new reinforcement learning system lets large language models challenge and improve themselves using real-world data instead of curated training sets. Meta researchers have unveiled a new ...