GOOGLE DEEPMIND'S ONLINE RLHF WITH 10X DATA EFFICIENCY

Summary from the TLDR AI newsletter

An online RLHF algorithm incrementally updated reward and language models, achieving over 10x data efficiency gains using uncertainty modeling and information-directed exploration.