A training method that encourages models to respond symmetrically to paired prompts representing opposing perspectives.