Toward Optimal LLM Alignments Using Two-Player Games