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

Abstract: Modern LLMs like ChatGPT have shown remarkable performance on general language tasks but still struggle on complex reasoning tasks, which drives the research on cognitive behaviors of LLMs to explore human-like problem-solving strategies. Along this direction, one representative strategy is self-reflection, which asks an LLM to refine the solution with the feedback generated by itself iteratively. However, our study shows that such reflection-style methods suffer from the Degeneration-of-Thought (DoT) problem: once the LLM has established confidence in its solutions, it is unable to generate novel thoughts later through reflection even if its initial stance is incorrect. To address the DoT problem, we propose a Multi-Agent Debate (MAD) framework, in which multiple agents express their arguments in the state of "tit for tat" and a judge manages the debate process to obtain a final solution. Clearly, our MAD framework encourages divergent thinking in LLMs which would be helpful for tasks that require deep levels of contemplation. Experiment results on two challenging datasets, commonsense machine translation and counter-intuitive arithmetic reasoning, demonstrate the effectiveness of our MAD framework. Extensive analyses suggest that the adaptive break of debate and the modest level of "tit for tat" state are required for MAD to obtain good performance. Moreover, we find that LLMs might not be a fair judge if different LLMs are used for agents. Code is available at https://github.com/Skytliang/Multi-Agents-Debate.

• The paper highlights the Degeneration-of-Thought problem, where established model confidence hinders the generation of novel ideas.
• It proposes a Multi-Agent Debate framework that leverages diverse agent interactions and external feedback to foster divergent thinking.
• Experimental results on commonsense machine translation and counter-intuitive arithmetic reasoning show MAD outperforming baseline models such as GPT-4 and GPT-3.5-Turbo.

Encouraging Divergent Thinking in LLMs through Multi-Agent Debate

Introduction

The potential of LLMs like ChatGPT and GPT-4 to perform well on general language tasks is widely acknowledged; however, they face challenges when addressing complex reasoning tasks. This paper introduces the concept of the Degeneration-of-Thought (DoT) problem observed in self-reflection methods used by LLMs, where an established confidence in solutions impedes the generation of novel thoughts even if initial stances are incorrect. To tackle this problem, the authors propose a Multi-Agent Debate (MAD) framework. This framework embodies multiple agents in a debate format that encourages divergent thinking, thereby improving performance on tasks requiring deep contemplation.

Degeneration-of-Thought Problem

The DoT problem, as defined in the paper, occurs when LLMs become unable to generate novel thoughts after developing confidence in their answers. This issue is particularly pronounced in self-reflection techniques, which rely heavily on the LLM's ability to evaluate its own statements accurately. The paper identifies factors contributing to the DoT problem, such as biases learned from data during training, rigidity in beliefs, and limited external feedback.

Figure 1: Degeneration-of-thought with respect to the iteration of self-reflection, where the proposed multi-agent debate method alleviates the DoT problem by producing more divergent thoughts.

Multi-Agent Debate Framework

To overcome the DoT problem, the paper proposes the MAD framework, which structures debates between multiple agents in a "tit for tat" state, monitored by a judge. This setup allows LLMs to exchange diverse arguments, correcting distorted thinking and obtaining external feedback from other agents. The MAD framework aims to facilitate divergent thinking, which is essential for solving complex reasoning tasks.

Figure 2: Framework of Multi-Agent Debate, illustrating agents in debate with a judge managing the process.

Experimental Evaluation

The authors conducted experiments using two testbeds: commonsense machine translation (Common MT) and counter-intuitive arithmetic reasoning (Counter-Intuitive AR). These tasks require LLMs to exhibit deep contemplative abilities beyond superficial interpretations. The results indicate that MAD significantly outperforms baseline methods, including GPT-4, especially when employing agents based on GPT-3.5-Turbo.

Figure 3: Translation performance with respect to the iteration of debate, demonstrating MAD's effectiveness compared to self-reflection.

Analysis and Implications

The analysis reveals that the adaptive break strategy and a modest level of "tit for tat" state are critical for optimizing MAD's performance. Moreover, the inherent bias in LLMs suggests that LLMs may not serve as fair judges when differently trained models are used as agents. These findings highlight the importance of multi-agent interactions and external feedback to ameliorate DoT and enhance reasoning capabilities in LLMs.

Figure 4: Translation performance with respect to the level of "tit for tat" state, highlighting the necessity for disagreement in MAD.

Conclusion

The paper presents a novel approach to address the Degeneration-of-Thought problem in LLMs through the Multi-Agent Debate framework. MAD encourages divergent thinking by leveraging debate among agents, effectively improving performance on complex reasoning tasks. Future research may explore expanding the number of agents, integrating AI feedback mechanisms, and applying the framework to broader AI alignment challenges. This study advances understanding of cognitive processes in LLMs and offers insights into enhancing their reasoning abilities.