Subtle Errors Matter: Preference Learning via Error-injected Self-editing