Unlocking the Power of Multi-Agent LLM for Reasoning: From Lazy Agents to Deliberation