International Journal of Computer Science and Mobile Applications

Diversity of Thought Elicits Stronger Reasoning Capabilities in Multi-Agent Debate Frameworks

Mahmood Hegazy^* Department of Computer Science, University of Montreal Mila, Quebec AI Institute, Canada
^*Correspondence: Mahmood Hegazy, Department of Computer Science, University of Montreal Mila, Quebec AI Institute, Canada, Email:

Received: Oct 23, 2024, Manuscript No. IJCSMA-24-150834; Editor assigned: Oct 25, 2024, Pre QC No. IJCSMA-24-150834(PQ); Reviewed: Oct 31, 2024, QC No. IJCSMA-24-150834(Q); Revised: Nov 06, 2024, Manuscript No. IJCSMA-24-150834(R); Published: Nov 13, 2024

Introduction

In the dynamic realm of artificial intelligence, enhancing the reasoning abilities and factual accuracy of Large Language Models (LLMs) stands as a pivotal challenge. Central to this pursuit is the exploration of innovative methodologies that address existing shortcomings and chart new pathways for advancement. Our research aims to fortify the foundations of LLMs through the lens of multi-agent debate.

The key motivation behind this project is solving the issue of "hallucination" within language models, where plausible yet erroneous information is generated, undermining their reliability and trustworthiness. Inspired by the collaborative nature of human intellectual discourse, the methodology of multi-agent debate emerges as a promising solution to this problem. By harnessing the collective insights of multiple AI agents engaged in structured debate, we seek to not only mitigate hallucinations but also elevate the precision and reliability of LLM responses.

A glance at the landscape of current research reveals a series of endeavours aimed at fortifying the reasoning capabilities of LLMs. While recent iterations of LLMs, such as GPT-4, represent significant strides forward, concerns persist regarding their reasoning capabilities. In response to these limitations, advocates the transformative potential of multi-agent debate. Furthermore, advancements in agentic approaches such as MetaGPT, as proposed, Agent verse, offer diverse perspectives, delving into collaborative problem-solving and simulation of human behaviour, thus broadening the horizons of LLM research (Figure 1).

Figure 1: Diverse model debate performance across 4 rounds on the ASDiv benchmark.

Our work aims to mitigate serious LLM deficiencies in reasoning by offering a nuanced analysis of existing multi-agent debate frameworks and their effectiveness in enhancing reasoning in LLMs. Building upon the insight from previous studies; our approach emphasizes the importance of diversity in models and debate strategies (Diversity of Thought) DOC. By synthesizing findings from diverse benchmarks and methodologies, we provide a comprehensive understanding of the strengths and limitations of current frameworks [1-3].

To empirically validate the effectiveness of our approach, we performed comprehensive experiments utilizing both diverse and homogeneous sets of language models with varying capacities. These experiments were conducted on multiple mathematical reasoning benchmarks, including the challenging GSM-8K dataset and the recently introduced ASDiv benchmark, which assess the models’ ability to generate accurate and well-reasoned solutions to complex problems. Our results demonstrate that leveraging diversity of thought in multi-agent debate significantly enhances the reasoning capabilities of LLMs, outperforming even state-of-the-art models like GPT-4. These findings underscore the potential of diverse, cooperating agents in achieving emergent capabilities beyond individual models.

Related Work

The landscape of Large Language Models (LLMs) has witnessed remarkable advancements in recent years, exemplified by innovations such as GPT-4 (OpenAI 2023), however, despite these breakthroughs, a critical examination reveals significant concerns regarding the reasoning capabilities of LLMs, as highlighted. This recognition has motivated a series of research efforts aimed at enhancing the reasoning and problem-solving abilities of LLMs through various methodologies.

Approaches like chain-of-thought prompting, self-verification, and multi-agent debate, have been proposed to address this challenge. The multi-agent debate approach, inspired by "Society of Mind" theory, posits that diverse agents approaching a problem with different methods, purposes, knowledge representations, and result-production techniques can enhance factual accuracy through debate and communication. Similarly, introduced MetaGPT, a meta-programming framework designed to tackle logic inconsistencies and hallucination by incorporating Standardized Operating Procedures (SOPs) and structured communication within LLM-based multi-agent systems.

In parallel, efforts have been directed towards enhancing the generative capabilities of LLMs to simulate believable human behaviour, These endeavours, while successful in creating realistic simulations, have also prompted further exploration into refining retrieval modules and real time interactivity to mitigate instances of hallucination [4-6].

Furthermore, frameworks such as Agent verse prioritize collaborative problem-solving among autonomous agents, emulating human group dynamics to achieve superior performance across diverse tasks. This emphasis on collaborative reasoning sets a precedent for future developments aimed at refining the intricacies of LLMs and advancing their capabilities in various domains (Figure 2).

Figure 2: Diverse model debate performance across 4 rounds on the GSM-8K benchmark.

Collectively, these studies underscore the imperative to address the reasoning deficits of LLMs while also exploring avenues for enhancing their generative capabilities and facilitating collaborative problem solving. Our work builds upon these foundations, leveraging the power of multi agent debate and diversity of thought to push the boundaries of LLM reasoning and pave the way for more reliable and capable language models [7].

Methodology

Our multi-agent debate framework for enhancing the mathematical reasoning capabilities of LLMs broadened the scope of the framework, as introduced to ensure it was compatible with diverse models architectures. As illustrated in figure 3. It consists of the following key components:

Figure 3: Multi agent debate framework architecture.

• Question Encoding: The mathematical question or problem is provided as input to the system. Thisquestion serves as the starting point for the debate among the participating models.

• Debating Models: Three diverse language models - Model 1, 2, and 3 are employed as the debating agents.These models can be chosen to have different architectures. We utilized this architecture torun experiments with and without diverse models.

• Debate Rounds: The debating models engage in multiple rounds of debate, where each model generates aresponse to the question based on its own reasoning capabilities. In each round, the models take turnsproviding their responses.

• Response Summarization: After each round, the responses from all three debating models are passedthrough a fourth model- Model 4 (Response Summarizer). This model’s role is to analyse and summarizethe key arguments, reasoning steps, and conclusions presented by the debating models. The summarizedresponse captures the most salient and convincing points from the debate round. Our model of choice forresponse summarization was Gemini-Pro for all experiments.

• Iterative Refinement: The summarized response from Model 4 is then fed back as input to the debatingmodels for the next round of debate. This iterative process allows the models to build upon each other’sreasoning, identify and correct errors, and refine their arguments based on the collective insights generatedin previous rounds.

• Final Summary: After a predefined number of debate rounds (n), the final summarized response from thesummarizer model (Model 4) is considered as the output of the multi-agent debate framework. This finalsummary represents the consolidated reasoning and conclusion arrived at through the iterative debateprocess. Here we can extract what the mode of the answers of the 3 models was.

The debate framework leverages the diversity of the participating models to explore different reasoning paths, challenge assumptions, and arrive at more robust and accurate conclusions. By encouraging the models to critically examine and build upon each other’s arguments, the framework aims to mitigate the limitations of individual models and enhance the overall mathematical reasoning capabilities of the system [8-11].

Datasets

To empirically evaluate the effectiveness of our multi-agent debate framework, we employ a diverse set of benchmarks that assess various aspects of language understanding and mathematical 0 reasoning.

• GSM-8K (Cobbe et al. 2021): This benchmark comprises 8.5K linguistically diverse grade schoolmath word problems, making it ideal for evaluating multi-step mathematical reasoning.

• Academia Sinica Diverse MWP Dataset (ASDiv) (Miao et al. 2021): ASDiv features 2,306 diversemath word problems covering various language patterns and problem types encountered in elementaryschool. It includes annotations for problem type and grade level.

• MATH (Enderton 2001): This historically challenging dataset provides 12,500 competitionmathematics problems, each accompanied by step-by-step solutions. It facilitates the teaching ofanswer derivations and explanations.

Through extensive experiments, we analyse the impact of model diversity, debate round count, and model size on reasoning performance. The results provide insights into the optimal configuration of the multi-agent debate framework for achieving superior mathematical reasoning capabilities compared to individual models.

By leveraging these diverse datasets, we aim to comprehensively assess the effectiveness of our approach in enhancing the reasoning capabilities of LLMs across a wide range of problem types, difficulty levels, and language patterns. This rigorous evaluation enables us to draw meaningful conclusions about the potential of multi-agent debate in advancing the state-of-the-art in language understanding and mathematical reasoning.

Experiments

To validate the effectiveness of our multi-agent debate framework, we conducted a series of experiments using diverse and homogeneous sets of language models with varying capacities. These experiments were performed on multiple mathematical reasoning benchmarks, as outlined in section 3.1, to assess the models’ ability to generate accurate and well-reasoned solutions to complex problems. Our two main goals of this study were as follows:

• Explore the relationship between model capacity and reasoning performance in the context of multi-agentdebate (Figure 4).

Figure 4: Illustration of the debate procedure.

• Investigate the impact of model diversity on the reasoning performance of the multi agent debate framework [12].

Baseline

To establish a fair and consistent baseline for comparison, we begin by asking each agent to directly generate responses to the given prompts without engaging in debate. This initial response generation serves as round 0 of the debate process and allows us to analyse the performance of the individual models before the collaborative reasoning begins.

By evaluating the models’ standalone performance in round 0, we can effectively measure the impact of the subsequent debate rounds on the overall reasoning capabilities of the system. This baseline assessment is crucial for understanding the extent to which the multi-agent debate framework enhances the models’ ability to generate accurate and well-reasoned solutions.

To ensure the validity and reliability of our comparisons, we maintain a consistent experimental setup across all evaluations. We use identical starting prompts and language models for both the baseline and the multi-agent debate framework approaches. This consistency eliminates potential confounding factors and enables us to attribute any observed improvements in performance to the effectiveness of the debate process itself [13-15].

Evaluation Methods

To facilitate a comprehensive evaluation of our multi-agent debate framework, we designed a systematic approach to assess the model outputs at each round of the debate. We configured the model prompts to consistently provide a boxed answer at the conclusion of each response, ensuring a standardized format for evaluation.

Following each experiment, we generate a JSON file that encapsulates the model outputs at each round of the debate for every question in the dataset. This structured data allows for a granular analysis of the framework’s performance throughout the debate process. Our evaluation script iterates through the debate rounds, assessing the accuracy and quality of the framework’s responses at each stage. By comparing the model generated answers to the ground truth solutions provided in the benchmark datasets, we can quantitatively measure the effectiveness of the multi-agent debate approach in enhancing the reasoning capabilities of LLMs [16, 17].

Furthermore, we visualize the performance of the framework across rounds, as exemplified in figure 5. This visual representation provides valuable insights into the progression of reasoning quality as the debate unfolds, allowing us to identify patterns, convergence points, and potential limitations of the approach.

Figure 5: Debate framework performance across model scales 0.75B to 100B+ on GSM8K.

Effect of Model Capacity on Performance

While previous work explored the effectiveness of multi-agent debate by varying the number of agents and the number of debate rounds, they did not investigate the effect of model scale on the framework’s performance. Considering the findings of which demon started that Chain-of-Thought (COT) reasoning is an emergent ability that arises with increased model size; we recognized the importance of conducting a similar experiment for multi agent debate.

To address this gap in the literature, we performed a scaling experiment on the GSM-8K dataset to examine the relationship between model capacity and the performance of the multi-agent debate framework. We evaluated the framework’s performance with and without COT reasoning across a range of model sizes, from small-scale models to large-scale ones. The results of our experiment, as presented in figure 6. Revealed a surprising finding the Contrary to the expectation that multi-agent debate performance would emerge as a direct consequence of increasing model size, we observed similar performance gains across all model scales. This result suggests that the effectiveness of multi-agent debate in enhancing reasoning capabilities is not solely dependent on the model’s capacity [18].

Figure 6: 7B Diverse models debate performance across 4 rounds on GSM8K dataset.

Our findings have significant implications for the development and deployment of multi-agent debate systems. They indicate that even smaller-scale models can benefit from the collaborative reasoning process facilitated by the debate framework, without the need for resource-intensive large-scale models. This insight opens up new possibilities for implementing multi-agent debate in resource- constrained environments and facilitates the widespread adoption of this approach.

Furthermore, our experiment highlights the importance of considering factors beyond model size when designing and optimizing multi-agent debate systems. It suggests that the effectiveness of the debate process may be influenced by other aspects, such as the diversity of the participating models, the structure of the debate, and the quality of the response summarization.

Diversity of Thought

Building upon the previous experiment, we aimed to investigate the impact of model diversity on the performance of the multi-agent debate framework. To introduce diversity, we replicated the study using three models of similar capacity but featuring diverse model families.

We first evaluated the framework’s performance on the GSM8K dataset using Gemini-Pro, when tested individually, these models achieved accuracy rates of 78%, 64%, and 70%, respectively, on the benchmark. Remarkably, the accuracy of the framework improved significantly, rising from 78% to an impressive 91% after 4 rounds of debate. This substantial improvement outperforms GPT-4 and highlights the power of collaborative reasoning and the synergistic effect of diverse perspectives.

To further emphasize the importance of diversity, we compared these results to a homogeneous setup consisting of 3 Gemini Pro models. In the homogeneous case, the performance improved from 78% to 80% without Chain-of-Thought (COT) reasoning and to 82% with zero shot COT. While still an improvement, the gains in the homogeneous setup were notably lower than those observed in the diverse model configuration.

These findings strongly suggest that the models in the diverse setup greatly benefited from the debate process, leveraging the unique reasoning approaches of their counterparts to refine and enhance their responses. The synergistic effect of combining models with different architectures and capabilities underscores the crucial role of collective insight in boosting overall performance.

Our study thus highlights a key insight: within the multi-agent debate framework, diversity is a critical driver of success. By fostering collaboration among models with complementary strengths, the framework enables the emergence of novel reasoning patterns and more accurate solutions to complex problems.

Qualitative Result

Figure 4 provides an illustrative example of the dynamics that unfolded during the multi-agent debate experiment. In this particular instance, we observe that Mixtral, one of the participating models, initially maintained its original answer for the first two rounds of the debate. However, by the third round, Mixtral’s stance began to shift as it carefully considered the reasoning put forth by the other two models. This pivotal moment in the debate showcases the power of diverse perspectives in challenging and refining the models’ understanding of the problem at hand. Mixtral, before fully adapting its reasoning to align with the collective insights, took a step back to articulate the rationale behind its initial divergent answer. This act of self-explanation not only adds transparency to the debate process but also highlights the model’s ability to engage in metacognitive reflection. By acknowledging its initial reasoning and subsequently integrating the persuasive arguments presented by its counterparts, Mixtral demonstrates the capacity for growth and learning within the multi-agent debate framework. This qualitative analysis offers a glimpse into the rich interplay of ideas and the collaborative knowledge construction that emerges when diverse models engage in structured debate. More of this can be found in the appendix [19].

To further validate the effectiveness of the multi-agent debate framework with diverse models, we conducted additional experiments on the ASDiv and MATH benchmarks. For these experiments, we employed three diverse models: Gemini Flash 1.5, Gemini Pro, and GPT 3.5. On the ASDiv benchmark, the individual performance of these models was already impressive, with Gemini Flash 1.5, Gemini Pro, and GPT 3.5 achieving accuracy rates of 89%, 86%, and 81%, respectively. However, when these models were combined in our multi-agent debate framework, the results were even more remarkable. After 4 rounds of debate, the framework reached an accuracy of 94%, setting a new state of the art performance on the ASDiv benchmark, surpassing the previous record set, as shown in figure 7.

Figure 7: 2B Diverse models debate performance across 4 rounds on GSM8K dataset.

Similarly, on the challenging MATH benchmark, the individual models achieved accuracy rates of 55%, 32%, and 33% for Gemini Flash 1.5, Gemini Pro, and GPT 3.5, respectively. When engaged in multi-agent debate, the framework’s performance significantly improved. By the 4th round of debate, the framework outperformed both GPT-4 and Gemini Ultra by substantial margins of 24% and 14%, respectively, as observed in figure 8.

Figure 8: Diverse model debate performance across 4 rounds on the math benchmark.

These results further underscore the power of diverse models in the multi-agent debate framework. By leveraging the unique strengths and perspectives of each model, the framework is able to achieve remarkable performance gains, pushing the boundaries of what is possible on these challenging benchmarks.

Moreover, to investigate whether diversity of thought is an emergent ability that arises with model scale, we conducted the same diversity experiment on GSM-8K using smaller models. The results were quite notable. Whether using 7B models (Gemma 7B, Mistral 7B, and Llama 2 7B) or 2B models (Gemma 2B, Qwen 2B, and Rho 1B), diversity of thought consistently elicited enhanced reasoning capabilities among these smaller models. As shown in Figures 6 and 7, the 7B model framework achieved a significant performance increase of 17% by the 4th round, while the 2B model framework saw a 10% improvement. These findings challenge our initial hypothesis that the framework’s initial competency in the specific benchmark would be crucial for facilitating a productive debate. Instead, our results suggest that the critical requirement for an effective debate is the presence of diverse model architectures of similar capacity, which induces learning and enhances reasoning capabilities.

The consistent performance improvements observed across different model scales highlight the robustness and generalizability of the multi-agent debate framework. The effectiveness of the approach is not limited to large-scale models but extends to smaller models as well, provided that architectural diversity is maintained. This finding has significant implications for the development and deployment of multi-agent debate systems, as it demonstrates the potential for achieving enhanced reasoning capabilities even with resource-constrained models [20].

Conclusion

In this paper, we have presented a comprehensive investigation into the effectiveness of multi-agent debate in enhancing the reasoning capabilities and factual accuracy of Large Language Models (LLMs). By building upon the foundational work, we have developed an advanced framework that leverages the power of diverse models and iterative refinement to push the boundaries of what is possible in collaborative reasoning and problem solving.

Our experiments on a range of challenging benchmarks, including GSM-8K, ASDiv, and MATH, have consistently demonstrated the remarkable performance gains achieved by the multi-agent debate framework. The diversity of thought inherent in the framework, brought about by the inclusion of models with different architectures and capabilities, has proven to be a critical driver of success. Through the iterative process of debate, the models are able to learn from each other, refine their reasoning, and converge on more accurate and robust solutions.

The framework’s ability to outperform even state-of-the-art models like GPT-4 and set new records on established benchmarks underscores its potential to revolutionize the field of AI. By harnessing the collective intelligence of diverse models, we have shown that it is possible to achieve emergent capabilities that surpass those of individual models, even the most advanced ones.

Moreover, the consistent effectiveness of the multi-agent debate framework across different model scales and datasets highlights its robustness and generalizability. This versatility opens up exciting possibilities for the application of the framework in various domains, from education and research to industry and beyond.

As we look to the future, our findings suggest that the key to unlocking the full potential of AI lies in fostering collaboration and diversity. By embracing the power of multi-agent systems and encouraging the development of models with complementary strengths, we can continue to push the boundaries of what is possible and address increasingly complex challenges.

In conclusion, our work represents a significant step forward in the quest to enhance the reasoning capabilities and factual accuracy of LLMs. The multi-agent debate framework, with its emphasis on diversity and iterative refinement, offers a promising path towards the development of more reliable, trustworthy, and capable AI systems. As we continue to explore and refine this approach, we invite the research community to build upon our findings and join us in shaping the future of collaborative agentic AI.

References

1. Arkoudas, Konstantine. "GPT-4 can't reason." 2308.03762 (2023).

   [Google Scholar] [Crossref]

2. Bubeck, Sébastien, et al. "Sparks of artificial general intelligence: Early experiments with gpt-4."2303.12712 (2023).

   [Google Scholar] [Crossref]

3. Chen, Weize, et al. "Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors in agents." 2308.10848 2.4 (2023): 6.

   [Google Scholar]

4. Chowdhery, Aakanksha, et al. "Palm: Scaling language modeling with pathways." J Mach Learn Res. 24.240 (2023): 100-113.

   [Google Scholar]

5. Cobbe, Karl, et al. "Training verifiers to solve math word problems." 2110.14168 (2021).

   [Google Scholar] [Crossref]

6. Du, Yilun, et al. "Improving factuality and reasoning in language models through multiagent debate."2305.14325 (2023).

   [Google Scholar] [Crossref]

7. Enderton, Herbert B. A mathematical introduction to logic. Elsevier, 2001.

   [Google Scholar]

8. Jiang, Albert Q., et al. "Mixtral of experts." 2401.04088 (2024).

   [Google Scholar] [Crossref]

9. Hong, Sirui, et al. "Metagpt: Meta programming for multi-agent collaborative framework." 2308.00352 (2023).

   [Google Scholar] [Crossref]

10. Kim, J. B., et al. "ATHENA: Mathematical reasoning with thought expansion." 2311.01036 (2023).

    [Google Scholar] [Crossref]

11. Miao, Shen-Yun,et al. "A diverse corpus for evaluating and developing English math word problem solvers."  2106.15772 (2021).

    [Google Scholar] [Crossref]

12. Minsky, Marvin. Society of mind. Simon Schuster, 1988.

    [Google Scholar]

13. Achiam, Josh, et al. "Gpt-4 technical report." 2303.08774 (2023).

    [Google Scholar] [Crossref]

14. Park, Joon Sung, et al. "Generative agents: Interactive simulacra of human behavior." Proc. 36th Annu Acm Symp User Interface Softw Technol. 2023.

    [Google Scholar] [Crossref]

15. Touvron, Hugo, et al. "Llama: Open and efficient foundation language models."2302.13971 (2023).

    [Google Scholar] [Crossref]

16. Wang, Guanzhi, et al. "Voyager: An open-ended embodied agent with large language models."2305.16291 (2023).

    [Google Scholar] [Crossref]

17. Wei, Jason, et al. "Chain-of-thought prompting elicits reasoning in large language models." Adv Neural Inf Process Syst. 35 (2022): 24824-24837.

    [Google Scholar]

18. Weng, Yixuan, et al. "Large language models are better reasoners with self-verification." 2212.09561 (2022).

    [Google Scholar] [Crossref]

19. Yao, Shunyu, et al. "React: Synergizing reasoning and acting in language models." 2210.03629 (2022).

    [Google Scholar] [Crossref]

20. Zhang, Peiyuan, et al. "Tinyllama: An open-source small language model." 2401.02385 (2024).

    [Google Scholar] [Crossref]

Appendix/supplemental material

This appendix presents a deeper dive into the debates, providing supplementary analysis and visual representations to enhance understanding.

Supplementary Analysis. Teacher-Student Behaviour Appearing in Framework with Diverse Capacities

Given the previous result of diversity of thought, we sought to better understand what is truly happening in the multi agent debate framework. We tested the framework while allowing one of the 3 models to be of higher capacity to observe if the smaller models can be taught at faster pace. The first experiment was done by using 2 small models (Gemma 7B and Qwen1.5 7B) and a 3^rd of higher capacity; Mixtral 7Bx8. The final outcomes are presented in figure a, where the effectiveness of the multi agent debate framework is showcased. Remarkably, without any rounds of debate and simply by adopting the summary answer of the three models, we achieve the best performance compared to the individual models at 69% accuracy. However, surprisingly, the framework’s performance does degrade to 66% as the models debate. Looking at the performance of individual models in a vacuum we can start to understand why. The 2 models that start at sub-30% accuracy, Gemma7B and Qwen1.5 7B, mark some impressive performance gains, with Gemma7B in particular punching above its weight and getting more than half the questions correctly by the 3^rd round. However, Mixtral, which notably was performing very well in this benchmark as expected et al. [2024], decreased its performance by more than 20% as soon as debate started but started to regain performance on the 2^nd round. One hypothesis for the observed behaviour suggests that the superior model, Mixtral, functions akin to a teacher for the two lesser models. These models learn and adapt throughout the debate process. Meanwhile, the "teacher" model’s performance diminishes as it becomes influenced by the lesser models. To confirm this hypothesis, a similar experiment was made where Gemma7B was made to be the "teacher" as it was coupled with smaller inferior models; Gemma2B and TinyLlama (1.1B parameter Llama-based architecture model introduced by Zhang et al. [2024]). The final outcomes are presented in figure b. Here we observe similar patterns to what we observed in figure a. and confirm our hypothesis. While the, now "teacher", Gemma7B decreases its performance as it is influenced by the 2 less capable models, the other models increase their performance as debate continues. With this finding we can adjust our previous theory to be that for a debate to be effective, the crucial requirement is that the participating models must have diverse architectures of similar sizes.

Figure : Illustration of teacher, student behaviour in debate framework on the GSM8K benchmark. a) Gemma 7B, Qwen1.5 7B, b) Mixtral 7Bx8 Gemma 7B, Gemma 2B and TinyLllama