Rethinking the Bounds of LLM Reasoning: Are Multi-Agent Discussions the Key?