More is more: Addition bias in large language models