“Scaling up language model size can worsen in-context reward hacking (ICRH).”