How are individual population groups affected? We find calibration tends to be more effective in improving alignment with groups for which the original LLM distributions are least aligned. E.g., for…