arxiv Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate