Free-MAD: Consensus-Free Multi-Agent Debate

Abstract: Multi-agent debate (MAD) is an emerging approach to improving the reasoning capabilities of LLMs. Existing MAD methods rely on multiple rounds of interaction among agents to reach consensus, and the final output is selected by majority voting in the last round. However, this consensus-based design faces several limitations. First, multiple rounds of communication increases token overhead and limits scalability. Second, due to the inherent conformity of LLMs, agents that initially produce correct responses may be influenced by incorrect ones during the debate process, causing error propagation. Third, majority voting introduces randomness and unfairness in the decision-making phase, and can degrade the reasoning performance. To address these issues, we propose \textsc{Free-MAD}, a novel MAD framework that eliminates the need for consensus among agents. \textsc{Free-MAD} introduces a novel score-based decision mechanism that evaluates the entire debate trajectory rather than relying on the last round only. This mechanism tracks how each agent's reasoning evolves, enabling more accurate and fair outcomes. In addition, \textsc{Free-MAD} reconstructs the debate phase by introducing anti-conformity, a mechanism that enables agents to mitigate excessive influence from the majority. Experiments on eight benchmark datasets demonstrate that \textsc{Free-MAD} significantly improves reasoning performance while requiring only a single-round debate and thus reducing token costs. We also show that compared to existing MAD approaches, \textsc{Free-MAD} exhibits improved robustness in real-world attack scenarios.

• The paper introduces a novel consensus-free MAD framework that aggregates full debate trajectories through a score-based decision mechanism.
• It employs an anti-conformity protocol that encourages agents to critically assess peer outputs, mitigating conformity bias and error propagation.
• Empirical results demonstrate up to 16% accuracy improvement and enhanced token efficiency, proving Free-MAD's scalability and robustness under adversarial conditions.

Free-MAD: A Consensus-Free Multi-Agent Debate Framework for LLM Reasoning

Introduction and Motivation

Multi-agent debate (MAD) frameworks have become prominent for enhancing LLM reasoning by leveraging multi-round agent interactions. Conventional MAD paradigms primarily use consensus-driven protocols, selecting the final output via majority voting after several debate rounds. However, such consensus-centric mechanisms are fundamentally limited—token inefficiency restricts scalability, conformity within LLMs causes error propagation, and majority voting can randomly penalize correct minority answers, undermining both accuracy and fairness. Evidence shows that final majority-chosen answers can perform worse than initial agent responses, illustrating the inadequacy of consensus-based protocols.

Figure 1: Existing MAD approaches may obtain final answers that are even less accurate than the initial ones.

The Free-MAD framework eliminates these constraints by proposing a consensus-free MAD approach with two central contributions: (1) a score-based decision mechanism evaluating the full debate trajectory—not just the last round, and (2) a debate protocol integrating anti-conformity (critical reasoning against the majority), thereby directly mitigating error propagation and improving agent independence.

Formalization and Methodology

MAD Protocol Decomposition

Free-MAD formally decomposes multi-agent debate into two stages:

• Debate Stage: $N$ agents iteratively exchange and update responses to a user query $q$, guided by a prompt $p$, for $R$ rounds, maintaining a matrix of all intermediate outcomes.
• Decision Stage: Unlike majority voting on the final round, Free-MAD uses a score-based aggregation over every agent’s trajectory. Each agent’s change in answer—especially critical shifts from previously held, now-discarded beliefs—is interpreted as evidence for increased correctness, weighted inversely by round number to minimize conformity bias.

This mechanism is robust to the “Silent Agreement” problem, where agents silently conform to a flawed majority and disables the dominance of random selection in case of diverse outputs.

Anti-Conformity Debate Protocol

Standard prompts induce agent conformity, favoring majority-aligned answers regardless of their correctness, which propagates errors. Free-MAD introduces a structured anti-conformity prompt (CoT-based), actively encouraging agents to find and describe flaws in their peers’ outputs, only changing their own stance upon rigorous identification of errors, not on majority presence. This balances independent reasoning with information assimilation from others, directly suppressing the default LLM conformity parameter in debate dynamics.

Score-Based Collective Decision

The protocol’s scoring schema explicitly tracks agent response shifts, penalizing abandoned answers while rewarding maintained or newly adopted ones. Random selection among equally scored candidates ensures mathematical robustness to attack scenarios. Token overhead is minimized as the protocol achieves high accuracy with even a single debate round ($R=1$), conferring major scalability advantages for practical deployment.

Empirical Evaluation

A comprehensive empirical study was conducted across eight benchmarks, spanning mathematical, logical, and knowledge-based reasoning. Free-MAD demonstrates superior reasoning accuracy, resiliency to adversarial communication attacks, and strong scalability at minimal token cost compared to consensus-driven baselines.

Figure 2: Comprehensive comparative experimental results for MAD frameworks across multiple benchmarks.

Key quantitative results:

• Accuracy Improvement: Free-MAD-n (anti-conformity) achieves 64.43% accuracy ($\uparrow$16% over baseline) with $R=1$, outperforming majority-voting variants even with multiple rounds.
• Token Efficiency: Free-MAD maintains baseline-equivalent or better accuracy while consuming fewer tokens and reducing execution latency, as consensus is not required and fewer rounds suffice.
• Robustness: Free-MAD exhibits minimal degradation under communication attacks, unlike consensus-based baselines that suffer up to 20% drops in accuracy.

Figure 3: Experimental results when $R=1$.

Scalability analysis aligns with theoretical expectations: accuracy is decoupled from the round count $R$, allowing parallel, efficient deployment. The system remains robust even when a significant subset of agents are compromised or blocked from the debate.

Comparative Analysis and Practical Implications

Free-MAD overcomes the key vulnerabilities of traditional MAD frameworks:

• Byzantine Robustness: Score-based decision landscape mitigates the impact of compromised agents, unlike majority voting or judge-based methods that are prone to failures under adversarial prompt injections.
• Fairness: No agent is privileged; all participate equally and independently, avoiding judge or role-based biases.
• Flexibility: The framework natively supports switching between conformity and anti-conformity debate modes, tunable by application context.

Theoretical modeling supports the system’s adaptability to heterogeneous agent populations, increased problem difficulty, and direct compatibility with sparse topologies or hybrid protocols.

Future Directions

Potential future research includes optimizing weight configurations within the scoring schema for task- and domain-specific robustness, expanding agent heterogeneity, and evaluating MAD resilience against a larger suite of adversarial attacks. Enhanced coverage of reasoning LLMs and more complex benchmarks will further validate generality and identify edge cases where consensus-free protocols are most effective.

Conclusion

Free-MAD presents an authoritative shift in MAD protocol design, demonstrating that consensus is not necessary for optimal reasoning outcomes in LLM multi-agent frameworks. By leveraging comprehensive debate trajectories and rigorous anti-conformity logic, Free-MAD attains superior accuracy, fairness, scalability, and resilience, offering a robust foundation for both practical deployment and further academic exploration of collaborative LLM reasoning architectures.