This paper reevaluates the claim that multi-agent discussions enhance the reasoning abilities of Large Language Models (LLMs) through systematic experiments. The authors propose a novel group discussion framework called Conquer-and-Merge Discussion (CMD), which simulates human group discussions. They find that a single LLM with strong prompts can achieve performance comparable to the best existing multi-agent discussion approaches on various reasoning tasks and LLMs. Multi-agent discussions outperform single agents only when no demonstrations are provided. The study also identifies two common interaction mechanisms in LLMs during discussions: judge mistakes and wrong answer propagation. Additionally, the results show that agents powered by weaker LLMs can improve their performance when interacting with stronger LLMs. The paper concludes with insights into when to use multi-agent discussions and highlights the importance of prompt engineering and the role of demonstrations in enhancing reasoning capabilities.This paper reevaluates the claim that multi-agent discussions enhance the reasoning abilities of Large Language Models (LLMs) through systematic experiments. The authors propose a novel group discussion framework called Conquer-and-Merge Discussion (CMD), which simulates human group discussions. They find that a single LLM with strong prompts can achieve performance comparable to the best existing multi-agent discussion approaches on various reasoning tasks and LLMs. Multi-agent discussions outperform single agents only when no demonstrations are provided. The study also identifies two common interaction mechanisms in LLMs during discussions: judge mistakes and wrong answer propagation. Additionally, the results show that agents powered by weaker LLMs can improve their performance when interacting with stronger LLMs. The paper concludes with insights into when to use multi-agent discussions and highlights the importance of prompt engineering and the role of demonstrations in enhancing reasoning capabilities.