Abstract
Multi-Agent Debate (MAD) has emerged as a promising paradigm for improving the performance of large language models through collaborative reasoning. Despite recent advances, the key factors driving MAD's effectiveness remain unclear. In this work, we disentangle MAD into two key components—Majority Voting and inter-agent Debate—and assess their respective contributions. Through extensive experiments across seven NLP benchmarks, we find that Majority Voting alone accounts for most of the performance gains typically attributed to MAD. To explain this, we propose a theoretical framework that models debate as a stochastic process. We prove that it induces a martingale over agents' belief trajectories, implying that debate alone does not improve expected correctness. Guided by these insights, we demonstrate that targeted interventions, by biasing the belief update toward correction, can meaningfully enhance debate effectiveness. Overall, our findings suggest that while MAD has potential, simple ensembling methods remain strong and more reliable alternatives in many practical settings.
Introduction
Multi-Agent Debate (MAD) has gained significant traction as a method to enhance the reasoning capabilities of Large Language Models (LLMs). The core intuition is that by allowing multiple agents to deliberate, critique, and refine their answers over several rounds, the group will eventually converge on a more accurate solution than any single agent could produce alone. However, this iterative process is computationally expensive and architecturally complex.
This research applies "Occam’s Razor" to the MAD paradigm, asking a fundamental question: Is the iterative "debate" actually improving reasoning, or are the observed gains simply the result of ensembling multiple model outputs? By disentangling the multi-agent component (having many outputs) from the debate component (the iterative interaction), we evaluate whether the complexity of MAD is truly justified over simpler alternatives like Majority Voting.
Theoretical Framework: Debate as a Martingale
To understand the dynamics of agent interaction, we introduce a theoretical framework using the Dirichlet-Compound-Multinomial (DCM) distribution. In this model, an LLM is viewed as having an internal "belief" (represented by a Dirichlet distribution) and a stochastic output process (Multinomial).
Crucially, we prove that under standard Bayesian updates, the debate process is a martingale. In probability theory, a martingale is a sequence of random variables where the future expectation is equal to the current value. In the context of MAD, this implies that while an individual agent's belief might fluctuate during a conversation, the expected belief of the group in the correct answer remains constant over time. Essentially, debate does not create new information; it merely reshuffles the knowledge already present in the initial pool of agents. Without an external signal of truth, agents are just as likely to be swayed by a wrong peer as they are to be corrected by a right one.
Experimental Methodology
We evaluated the effectiveness of MAD against Majority Voting across seven diverse NLP benchmarks:
- Reasoning & Math: Arithmetic and GSM8K.
- Knowledge & Logic: MMLU (Medicine and Logic subsets) and CommonsenseQA.
- Language Inference: HellaSwag.
- Alignment: HH-RLHF (human preference labeling).
- Open-Ended Tasks: CNN/DailyMail (summarization).
We compared several MAD architectures—including Decentralized (all-to-all communication), Sparse (limited communication), and Centralized (mediator-led)—against a Majority Voting baseline where agents provide independent answers and the most frequent response is selected.
Key Findings
1. The Ensembling Illusion
Our empirical results show that Majority Voting performed on par with or better than MAD across almost all benchmarks. For instance, on the GSM8K math benchmark using Qwen2.5-7B, Majority Voting achieved 94.0% accuracy, while the best-performing MAD variant reached only 90.3%. This suggests that most performance gains previously attributed to "deliberation" are actually the result of the "wisdom of the crowd" effect inherent in ensembling multiple independent outputs.
2. Flat Accuracy Trajectories
Consistent with our martingale proof, empirical plots of agent accuracy over five rounds of debate show nearly flat trajectories. The group does not systematically become "smarter" as the debate progresses; the average correctness level remains stagnant. This confirms that interaction alone, without a mechanism to favor correct signals, leads to a "random walk" rather than a convergence toward truth.
3. The Impact of Answer Extraction
We discovered that previous studies may have over-reported the benefits of MAD due to inconsistent answer-extraction methods. When we applied a standardized format (e.g., requiring a specific JSON key for the final answer) to single-agent baselines, the performance gap between simple voting and complex debate narrowed significantly.
4. Breaking the Martingale
To make debate effective, the process must be "biased" toward correction. We demonstrated that by intervening in the update process—such as making agents more likely to conform to a majority or "locking in" signals that appear correct—we could finally outperform simple voting. This suggests that the future of MAD lies not in more conversation, but in better signal preservation.
Conclusion
While Multi-Agent Debate is an intuitive concept, our study reveals that simple ensembling via Majority Voting is often a more reliable and efficient alternative. Debate, in its standard form, acts as a martingale process that aggregates but does not amplify correctness. For MAD to surpass simple voting, researchers must design protocols that actively favor truth-seeking biases over stochastic peer influence. Our findings suggest that in many practical settings, the most effective way to use multiple LLMs is to let them work independently and vote on the result.
Summary updated