#semi-online training

Meta and NYU developed a semi-online reinforcement learning method that balances offline and online training to enhance large language model alignment, boosting performance in both instruction-based and mathematical tasks.