According to StanfordAI Lab on X, Freeform Preference Learning uses natural language axes to learn conditional rewards and yield better robot policies.