Multi-Agent Debate

Multi-agent debate is an approach in the field of large language models (LLMs) where several interacting agents (instances of a language model) collaboratively discuss the solution to a given problem by exchanging arguments and potential answers. The goal of this process is to collectively produce the most correct and well-reasoned answer to the question posed. The approach is based on the concept of a "society of mind," where different models check and supplement each other's conclusions. Studies have shown that multi-agent discussion can significantly increase the accuracy and reliability of answers compared to single-agent generation: the final answer obtained after an agent debate is typically more factually accurate and performs better on tasks requiring reasoning. Specifically, this strategy has been observed to reduce the number of hallucinations (non-existent "facts") and increase success rates on complex test tasks.

The idea of using multiple AIs for debate originates from work in artificial intelligence safety. In 2018, a group of OpenAI researchers (G. Irving, P. Christiano, D. Amodei) proposed the concept of AI safety via debate—training agents through adversarial debates in which two model opponents take turns presenting short arguments, and a human judge decides which one provided more truthful and useful information. It was hypothesized that, with an optimal strategy, such debates would allow an AI to answer extremely complex questions, requiring the judge only to assess the plausibility of the arguments. In subsequent years, with the emergence of powerful LLMs, the principle of debate between models began to be applied directly to improve the quality of the models' own answers—without mandatory human participation and with automated selection of the best solution. Modern multi-agent LLM systems use dialogue between copies of the same model or different models to correct each other's errors and jointly arrive at a more well-founded result.

The Multi-Agent Debate Procedure

In a multi-agent debate scenario, several agent models work on the same task in parallel. Typically, each agent is first given the initial question or problem, after which each agent independently generates its own answer. This is followed by a series of communication rounds between the agents: in each round, all participants share their current solutions, and each agent receives the others' answers as additional context, based on which it refines or improves its answer in the next round. This cycle continues for several iterations (usually a fixed number of rounds or until explicit agreement is reached), after which the process stops and a final answer is produced. Debates simulate human discussion, allowing models to critique each other's answers and combine their reasoning skills to enhance the quality of the solution. For example, Yilun Du and colleagues (MIT and Google Brain) used 3 instances of a language model in their experiments, which discussed a problem for 2 rounds (more rounds were limited due to time and computational costs); it was shown that even with such a limited dialogue, the final answers were noticeably better, and increasing the number of agents or rounds continued to improve accuracy (albeit with diminishing returns).

The multi-agent debate procedure is fully implemented at the inference stage using special prompts to organize the dialogue between already trained models. This means that the method does not require fine-tuning the LLMs themselves and can be applied even to "black box" models—it is sufficient to have access to the models' text generation capabilities and to coordinate their communication according to a predefined template.

Different approaches are used to determine the final answer after several rounds. One of the simplest mechanisms is voting: agents can independently propose their final solutions at the end, after which the option supported by the majority (or, for example, the most frequently occurring answer) is chosen. Another approach is to require consensus, meaning the discussion continues until all models arrive at the same answer. Finally, a separate judge agent may be involved: either a separate neural network trained to evaluate answers, or one of the agents assigned the role of an arbiter. The judge observes the discussion and selects whose argument was the most convincing or correct. The choice of decision-making mechanism affects the system's characteristics: for instance, voting or consensus is simple to implement but can lock in group errors, whereas a judge (especially one trained to identify the correct answer) is theoretically capable of isolating the right solution even amid contradictions between agents. However, the judge-based approach also has its difficulties—for example, if the judge is the same model as the participants, it may be unintentionally biased towards the familiar argumentation style of one of the agents.

Agent and Communication Configurations

A multi-agent system with LLMs can vary in the composition and interaction style of its agents. A homogeneous configuration assumes that all agents are copies of the same model (or models of a similar capability level), whereas a heterogeneous one includes models of different types or sizes. In the homogeneous case, all participants have comparable abilities, and their disagreements arise only from the stochastic nature of answer generation or different initial conditions (e.g., variations in prompts). In the heterogeneous approach, strong and weak models can be used simultaneously, potentially allowing some agents to compensate for the shortcomings of others. For example, research shows that interaction between different LLMs leads to weaker models improving their solutions by receiving feedback from stronger ones. A telling example is a joint debate between the ChatGPT (GPT-4) and Google Bard language models while solving a math word problem: each of these models gave an incorrect answer individually, but during the discussion, they managed to point out each other's errors and ultimately agree on the correct solution, leveraging the strengths of each. At the same time, heterogeneity also carries risks: a significant imbalance in capabilities can lead to the dominance of one model, and if a majority of agents share a common misconception or bias, the debate may quickly converge to a unified but incorrect answer—a phenomenon known as the "echo chamber" effect. Theoretical analysis (Estornell & Liu, NeurIPS 2024) has shown that with very similar models, a debate can stagnate into a static dynamic where all participants repeat the majority opinion, even if it is based on a shared error in their data. Therefore, in heterogeneous systems, careful agent selection is crucial—for example, by choosing models with comparable knowledge levels so that none dominates or misleads the others.

Another aspect is the communication structure between agents. Basic implementations use a fully connected topology of communication: in each round, every agent receives the answers of all others. This "all-to-all" exchange maximizes available information but creates significant overhead—the context volume grows proportionally with the number of agents, making computations more demanding. An alternative is a sparse topology, which limits with whom each agent directly exchanges data. For example, agents can be arranged in a network graph (a ring, a tree, etc.), where each agent receives answers only from its neighbors. A study by Google (Li et al., 2024) found that limiting the connectivity of the agent network can significantly reduce generation costs without degrading, and sometimes even improving, the quality of the solution compared to a fully connected discussion. In experiments with GPT-3.5 and Mistral models, a sparse "neighborly" discussion scheme yielded the same or higher accuracy on tasks (including mathematics), while reducing the average number of context tokens per step by an order of magnitude. This result suggests that excessive message exchange is not always necessary—it is sufficient to properly organize key interactions between agents for them to arrive at the correct solution with lower costs.

Besides topology, different debate formats are possible. For example, agents can be assigned different roles: some act as "idea generators," others as "critics" or "verifiers" of solutions. This role-based approach aims to simulate a division of labor, where each agent specializes in a specific task (e.g., one proposes a hypothesis, a second checks facts, a third assesses logical consistency). Another variant is a round-robin discussion: agents speak sequentially rather than simultaneously, taking turns as the speaker and responder in a fixed order. This resembles formal debates where participants are given the floor according to a schedule, which can ensure equal participation of all agents. Yet another approach is the dynamic regulation of disagreement: the system can intentionally strengthen or weaken the degree of disagreement between agents' answers in each round. For instance, it can encourage answers to diverge as much as possible in the early stages (to cover different hypotheses) and to converge as the debate nears its conclusion. Such a mechanism was proposed in a paper by Chang (2024) to prevent premature agreement: it maintains a moderate level of contradiction between agents, stimulating the emergence of new arguments and a deeper discussion.

Advantages and Effectiveness of the Approach

Multi-agent debates have gained attention for their ability to enhance the performance of language models on complex tasks. A series of independent studies from 2023-2024 has confirmed that a group of interacting LLMs can outperform a single model working on the same task. In particular, improvements have been shown in domains requiring complex reasoning, from mathematical calculations to programming and text summarization. For instance, Yin et al. (2023), Chan et al. (2023), Chen et al. (2024), and others note that multi-agent systems consistently outperform single LLMs in arithmetic problems, code generation, and even in creating document summaries. The reason is the diversity of perspectives: each agent can notice details or errors missed by others and provide feedback to its peers. Mutual criticism and the exchange of different hypotheses lead to a more comprehensive examination of the problem, making the final answer more accurate and reliable.

For example, researchers from MIT and Google Brain, led by Yilun Du, presented a paper at ICML 2024, "Improving factuality and reasoning in language models through multiagent debate," which demonstrated a significant improvement in the quality of solutions when debates between three instances of a model were added. According to their results, the multi-agent discussion procedure led to higher scores on several tasks compared to the standard single use of the same model: the accuracy in solving mathematical and strategic problems increased, and the number of factual errors decreased. Specifically, the multi-agent approach improved the model's performance on tests of mathematical reasoning, fact-checking, and even tasks requiring strategic planning. The authors note that "the final answer generated after such a multi-round discussion is both more factually correct and more successful in solving reasoning problems". Below is an illustration comparing the accuracy of a model on various tasks when used alone versus with multi-agent debates.

Comparison of accuracy on several tasks for single-user generation (blue) and for multi-agent debate mode (red). The multi-agent debate approach demonstrates higher accuracy across various domains, including factual questions (biographies), the MMLU knowledge test, checking the correctness of chess moves, solving arithmetic expressions, school-level math word problems (GSM8K), and finding the optimal chess move. According to the graph, debates particularly enhance the model's abilities in complex strategic tasks (e.g., finding the optimal move in chess) and significantly reduce the rate of errors in mathematical calculations and factual knowledge questions.

Another advantage of the multi-agent approach is overcoming the limitations of single-agent self-correction. Single LLMs often use the self-reflection technique, where the model itself evaluates and corrects its initial answer. However, this method has been found to be prone to the "degeneration-of-thought" problem: if the model is confident in its initial answer, it does not generate fundamentally new ideas during self-review, even if the original solution is wrong. In other words, the model tends to get stuck on its initial solution, rejecting alternatives. Multi-agent debates help to mitigate this effect: several equal agents can initially propose different hypotheses and then sequentially challenge each other's arguments, stimulating the search for unconventional lines of thought. Tian Liang and colleagues (EMNLP 2024) named their multi-agent scheme MAD (Multi-Agent Debate) and showed that it indeed encourages divergent thinking in models and improves results on tasks requiring deep problem-solving. In their implementation, several agents argue on a "tit-for-tat" basis (each taking turns to oppose another's arguments), while a supplementary judge oversees the process, managing the discussion and selecting the final solution. The experiments by Liang et al. demonstrated the effectiveness of this approach on complex test sets—in commonsense translation tasks (translating sentences that require an understanding of implicit common sense) and in counter-intuitive arithmetic (math puzzles with seemingly illogical conditions), the multi-agent discussion yielded more correct answers than standard methods. The analysis also revealed that for the best results, debates should be adaptively terminated to avoid excessive length, and only a moderate level of conflict should be maintained between agents—behavior that is too aggressive or, conversely, too agreeable degrades the outcome.

The multi-agent approach has proven useful not only for typical question-answering tasks. It is being applied in other areas, such as achieving safer and more aligned behavior in models. Some studies use agent debates for moderation and policy-making tasks: several LLMs can discuss whether a given response is acceptable according to ethical norms, thereby providing feedback to each other during reinforcement learning. It has been noted that debates can generate more nuanced and well-reasoned evaluation signals, which help in fine-tuning models for safety and helpfulness. There have also been attempts to extend this to multimodal tasks—for example, where some agents describe an image while others check if the description matches the picture. A paper from Google (2024) showed the success of such an extension: the multimodal approach improved results in both text-only tasks and in multimodal image understanding, demonstrating the versatility of the "society of mind" concept. Interestingly, interaction within a debate can elevate the performance of weaker models, as mentioned earlier. For instance, when LLMs of varying capabilities participate in a common discussion, "weaker models gradually improve by adopting successful strategies from stronger ones". Thus, a multi-agent system not only solves the task at hand but also serves as a mechanism for collective learning among models.

Limitations and Open Problems

Despite its significant advantages, multi-agent debate faces a number of challenges and limitations. One of the main ones is the high resource consumption of this approach. Organizing a discussion requires repeatedly calling the text generation function of large models: if n agents participate in T rounds, the total number of calls to the LLM increases by a factor of n x T compared to a single answer. Moreover, in each round, the model must process not only the original question but also all the remarks from previous rounds (the answers of all agents). Thus, as the number of agents and rounds increases, the volume of the context input grows exponentially, leading to context explosion—exceeding the context window and increasing processing costs. Experiments show that adding even 2-3 rounds of discussion significantly increases the total number of context tokens the model must read, and consequently, the response time. While solution quality theoretically improves with more iterations, many studies note diminishing returns after a few rounds: often, the maximum effect is achieved in the second or third round, after which further discussions can lead to the repetition of the same arguments or even a decrease in accuracy due to context overload. For example, He et al. (2023) showed an increase in accuracy only up to the 2nd round of debate, followed by a decline, and similarly, Liu, Li, and colleagues (2024) report a peak in quality at ~4 rounds, after which additional cycles are detrimental. Therefore, determining the optimal duration for a debate is a non-trivial task: a discussion that is too short may not unlock the full potential of collective intelligence, while one that is too long can cause information noise and context overload.

Another problem is the risk of group consensus on an incorrect answer. If all agents have similar experiences and are mistakenly confident in a certain fact, they can reinforce each other's delusion. This leads to an echo chamber effect: during the debate, the models reach a consensus, not because they have found the truth, but as a result of confirming their initial shared bias. Theoretical results (Estornell & Liu, 2024) indicate that with identical models, debates can descend into stagnation, repeating the majority opinion without generating new ideas. This is particularly dangerous when the majority shares a common error, for instance, one embedded in the training data—in which case the entire discussion will lead to an incorrect outcome. To overcome this problem, special intervention methods (diversity-pruning) are proposed: in each round, overly similar answers are algorithmically pruned, encouraging agents to generate diverse options with maximum information entropy. This reduces the likelihood that all answers will be variations of the same error. Another technique is misconception refutation: the system attempts to automatically identify the common assumptions of the agents and deliberately challenges those that may be false. The work by Estornell & Liu proposed a set of three such interventions—in addition to the ones mentioned, also quality-pruning (selecting the most relevant and high-quality arguments at each step)—and showed that their combination significantly improves the effectiveness of debates and prevents the tendency towards an echo chamber.

Finally, it should be noted that the stability and predictability of multi-agent discussions are still far from ideal. In some experiments, debates have led to unstable results—different runs of the same discussion could converge to different answers, or the collective answer could be worse than that of a single model without a debate. Wang et al. (2024) and Smit et al. (2023) independently reported cases where adding agents degraded performance, indicating a fine line between useful criticism and destructive arguments. Identifying the conditions under which the multi-agent approach is guaranteed to be beneficial remains a subject of research. Open questions include: how to automatically decide when to stop the debate and finalize the answer to capture the benefits without getting into an endless argument, and how to make the collective decision—whether through voting, consensus, or an external judge—most reliably for different types of tasks. The problem of safety and controllability of multi-agent systems is also acute: it is necessary to ensure that agents do not collaboratively generate undesirable or toxic content and do not amplify each other's harmful tendencies. These issues, especially those concerning safety and scalability, are recognized as current and complex. Modern reviews note that further research is needed to develop robust stopping criteria for discussions, evaluate the scalability of the approach as the number of agents and rounds increases, and implement methods to guarantee the reliability and faithfulness of the collectively obtained answer. Solving these challenges will help turn multi-agent debates into an even more powerful and versatile tool for creating smarter and safer artificial intelligence systems.

Links

• Improving Factuality and Reasoning in Language Models with Multiagent Debate — project page
• AI safety via debate — original OpenAI paper
• Improving Multi-Agent Debate with Sparse Communication Topology — paper on communication optimization
• Literature Review Of Multi-Agent Debate For Problem-Solving — literature review
• Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate — paper on MAD
• Improving Multi-Agent Debate with Contrastive Deliberation and Diversity-Promoting Interventions — NeurIPS 2024 paper

Literature

• Irving, G. et al. (2018). AI Safety via Debate. arXiv:1805.00899.
• Du, Y. et al. (2023). Improving Factuality and Reasoning in Language Models through Multiagent Debate. arXiv:2305.14325.
• Liang, T. et al. (2023). Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate. arXiv:2305.19118.
• Li, Y. et al. (2024). Improving Multi-Agent Debate with Sparse Communication Topology. arXiv:2406.11776.
• Guo, T. et al. (2024). Large Language Model based Multi-Agents: A Survey of Progress and Challenges. arXiv:2402.01680.
• Li, J. et al. (2024). More Agents Is All You Need. arXiv:2402.05120.
• Estornell, A.; Liu, Y. (2024). Multi-LLM Debate: Framework, Principals, and Interventions. NeurIPS 2024.
• Eo, S. et al. (2025). Debate Only When Necessary: Adaptive Multiagent Collaboration for Efficient LLM Reasoning. arXiv:2504.05047.
• Tillmann, A. (2025). Literature Review Of Multi-Agent Debate For Problem-Solving. arXiv:2506.00066.
• Cui, Y. et al. (2025). Efficient Leave-One-Out Approximation in LLM Multi-Agent Debate Based on Introspection. arXiv:2505.22192.
• La Malfa, E. et al. (2025). Large Language Models Miss the Multi-Agent Mark. arXiv:2505.21298.