---
title: 'Free-MAD: Consensus-Free LLM Debate'
url: https://www.emergentmind.com/topics/free-mad
type: topic
---

# Free-MAD: Consensus-Free LLM Debate

Free-MAD is a multi-agent debate framework for large language model reasoning that removes the requirement that agents reach consensus and replaces last-round majority voting with a deterministic, score-based evaluation of the entire debate trajectory. The framework uses a single-round debate, introduces anti-conformity to mitigate excessive majority influence, and targets three weaknesses of consensus-based MAD: multi-round token overhead, error propagation driven by LLM conformity, and the randomness and unfairness of majority voting when agents remain divided or converge to a wrong majority [2509.11035].

## 1. Conceptual basis and problem setting

Conventional multi-agent debate frameworks organize agents to debate over \(R\) rounds and then commit to the final answer by majority vote on the last round. In Free-MAD, that design is treated as suboptimal for three reasons. First, achieving consensus typically requires \(2\)–\(3\) rounds, so token cost grows rapidly with the number of agents \(N\) and rounds \(R\). Second, LLMs exhibit conformity: during debate, agents bias toward majority viewpoints even when the majority is incorrect. Third, majority voting introduces randomness and unfairness when the final round contains ties or all-distinct answers.

Free-MAD addresses these issues by using a single-round debate, excluding the initial independent response from \(R\), and by scoring all answers that appear throughout the debate rather than only the final round. The decision rule therefore tracks how each agent’s reasoning evolves across the trajectory. The framework has two variants. **Free-MAD-n** combines anti-conformity debate with score-based decision, while **Free-MAD-c** combines conformity-oriented debate with the same score-based decision rule. Both share the same trajectory-level decision mechanism, but differ in how the prompting modulates conformity pressure [2509.11035].

A central consequence is that the last round no longer has privileged status. An answer can be selected even if it is not the majority answer in the final round, provided that its cumulative trajectory score remains highest. This suggests a shift from consensus formation to trajectory-weighted evidence accumulation, although the underlying paper formulates this change operationally rather than as a separate theory of deliberation.

## 2. Debate protocol and anti-conformity formulation

Free-MAD decomposes MAD into a **Debate** stage and a **Decision** stage. Let \(N\) be the number of agents, \(q\) the query, and \(p\) the guiding prompt. At the initial step \(k=0\), each agent \(a_i\) independently samples an initial response \(r_i^0\). For debate rounds \(k=1,\ldots,R\), each agent updates its response given the shared context \(C^{(k-1)}\), with

\[
r_i^{k} \sim P_{a_i}(r \mid C^{(k-1)}, p).
\]

To separate independent reasoning from peer influence, the framework models

\[
P_{a_i}(r \mid C^{(k-1)}, p)
=
\frac{1}{Z}
\, P_{\text{in}}(r \mid q, p)\,
\exp\!\big(\beta(p)\, S_{\text{con}}(r, C^{(k-1)})\big),
\]

where \(P_{\text{in}}(r \mid q, p)\) captures the agent’s intrinsic reasoning ability given \(q\) and \(p\), \(S_{\text{con}}(r, C^{(k-1)})\) measures alignment of candidate \(r\) with peer utterances in the context, \(\beta(p)\) is a conformity parameter set by the prompt, and \(Z\) is a normalizing constant. When \(\beta(p) > 0\), the prompt encourages conformity; when \(\beta(p) < 0\), the prompt penalizes alignment and implements anti-conformity.

The anti-conformity mechanism is both probabilistic and prompt-based. Mathematically, \(\beta(p) < 0\) regularizes the response distribution toward \(P_{\text{in}}\). Operationally, the anti-conformity prompt instructs each agent to lay out initial logical steps and conclusion, analyze peers’ reasoning and identify concrete errors, compare against its own reasoning, revise only with explicit error justification, never adopt by majority, retain its conclusion if unsure, and independently identify errors each round without copying others. The paper characterizes this as a structured, critical CoT prompt that reduces blind herd behavior while still allowing justified revisions [2509.11035].

The debate trajectory of agent \(i\) is

\[
T_i = \{r_i^0, r_i^1, \ldots, r_i^R\},
\]

and Free-MAD evaluates the full trajectory matrix rather than only its last column. The framework’s headline design choice is that \(R=1\) is sufficient in the main setting: agents first produce independent answers, then engage in one round of interaction with peers’ reasoning traces, after which the score-based decision mechanism is applied.

## 3. Trajectory scoring and consensus-free decision

The decision mechanism maintains a score dictionary \(S\) keyed by distinct answers observed across agents and rounds. Let \(f_k\) be a round-dependent correction factor and let \(w_1,w_2,w_3,w_4\) be system-defined weights. Free-MAD sets

\[
f_k = (k+1)^{-1},
\]

so later rounds are downweighted to limit the influence of conformity that grows with more peer context.

For any answer \(a\), the trajectory score is

\[
S(a)
=
\sum_{k=0}^{R}\sum_{i=1}^{N}
f_k
\Big[
w_1 \mathbf{1}\{k=0\}\mathbf{1}\{r_i^0=a\}
+
w_3 \mathbf{1}\{k>0\}\mathbf{1}\{r_i^k=a\}\mathbf{1}\{r_i^{k-1}\neq a\}
+
w_4 \mathbf{1}\{k>0\}\mathbf{1}\{r_i^k=a\}\mathbf{1}\{r_i^{k-1}=a\}
-
w_2 \mathbf{1}\{k>0\}\mathbf{1}\{r_i^{k-1}=a\}\mathbf{1}\{r_i^k\neq a\}
\Big].
\]

The four terms have explicit interpretations. **Initial credit** \(w_1\) rewards independently generated answers before peer influence. **Adoption bonus** \(w_3\) rewards agents who switch to a new answer, signaling that they found better reasoning. **Stability bonus** \(w_4\) rewards agents who maintain an answer across rounds. **Abandon penalty** \(w_2\) penalizes answers that agents leave, signaling detected flaws. In experiments, the weights are initialized as

\[
\mathcal{W} = \{w_1,w_2,w_3,w_4\} = \{20,25,30,20\}.
\]

The final decision is

\[
MS = \operatorname{argmax}_a S(a),
\qquad
r_{\text{final}} \in
\begin{cases}
MS, & |MS|=1,\\
\text{random choice from } MS, & |MS|>1.
\end{cases}
\]

Accordingly, Free-MAD is deterministic except in rare ties across maximizers. The framework describes randomized tie-breaking as preserving theoretical robustness.

This scoring rule alters both fairness and failure modes. The paper states that evaluating full trajectories makes outcomes more fair and accurate: trajectory scoring reduces error propagation by preserving the influence of initial independent reasoning through \(w_1\), penalizes abandoned answers through \(w_2\), mitigates the randomness of majority voting in ties or all-distinct outcomes, and avoids an LLM-as-a-Judge hierarchy because agents are equal peers and the final score update is executed outside LLM generation [2509.11035].

## 4. Computational profile and empirical performance

Free-MAD measures token consumption as

\[
\text{TC} = \sum_{k=0}^{R}\sum_{i=1}^{N}\mathcal{T}_k^i,
\]

where \(\mathcal{T}_k^i\) is the number of output tokens generated by agent \(a_i\) at round \(k\). As a motivation baseline, prior MAD token complexity is summarized as

\[
O(N R^2 V + N^2 R V + N R |q|),
\]

where \(V\) is the per-agent token cost per round and \(|q|\) the query length. Free-MAD’s single-round debate reduces \(R\) to \(1\), excluding initial generation, and the paper reports that it achieves accuracy comparable to or higher than two-round baselines with almost no additional token cost beyond a single debate round.

The evaluation covers eight datasets: GSM-Ranges (Levels 4 and 6), AIME2024, AIME2025, MATH500, StrategyQA, MMLU Logical Fallacies, and AICrypto multiple-choice. Accuracy and token consumption are the primary metrics. For AIME2024 and AIME2025, the agent group uses \(N=3\), Qwen1.5-7B-Chat and DeepSeek-V3. For the other datasets, \(N=4\), Qwen1.5-7B-Chat and Qwen2.5-72B-Instruct are used. Temperature is set to default, tokens are counted via the DeepSeek-V3 tokenizer, full AIME sets are used, and the other datasets use \(80\) samples each.

| Variant | Debate mode | Decision rule |
|---|---|---|
| Free-MAD-n | Anti-conformity | Score-based |
| Free-MAD-c | Conformity-oriented | Score-based |
| Baseline 1 | Anti-conformity | Majority voting |
| Baseline 2 (SoM) | Conformity | Majority voting |

Under the single-round setting, average accuracy is \(64.43\%\) for Free-MAD-n, \(61.41\%\) for Free-MAD-c, \(55.73\%\) for Baseline 1, and \(54.06\%\) for Baseline 2. The paper states that Free-MAD-n improves by \(16\%\) and \(19\%\) over Baseline 1 and Baseline 2, respectively, while Free-MAD-c improves by \(10\%\) and \(14\%\). Under two rounds, the reported average accuracies over GSM-L4, GSM-L6, MATH500, Logical Fallacies, AICrypto, and StrategyQA are \(70.42\) for Free-MAD-n, \(69.17\) for Free-MAD-c, \(62.50\) for Baseline 1, and \(62.71\) for Baseline 2. The detailed two-round results include, for Free-MAD-n, GSM-L4 \(75.00\), GSM-L6 \(66.25\), MATH500 \(65.00\), Logical Fallacies \(71.25\), AICrypto \(76.25\), and StrategyQA \(68.75\) [2509.11035].

## 5. Robustness, fairness, and observed limitations

Free-MAD is also evaluated under a real-world attack scenario based on communication attacks, where compromised agents do not receive peers’ responses. This models unreliable networks and prompt-injection risks. The reported attack rate is \(50\%\), expressed as \(|\mathcal{V}|/|N| = 50\%\). Under this setting, Baseline 1 and Baseline 2 experience accuracy drops of up to \(20\%\), whereas Free-MAD maintains very high accuracy and in some cases slightly improves, because reduced exposure to incorrect peer content lowers error propagation. The paper attributes part of this robustness to the fact that the score mechanism executes deterministically outside LLM generation, making it immune to hallucinations and providing byzantine robustness.

The ablations separate the effects of anti-conformity and score-based decision. Score-based decision dominates majority voting across benchmarks, showing consistent gains in accuracy and fairness. Anti-conformity brings notable gains on hard math tasks, but weaker models may become too rigid, failing to switch to better reasoning lines in MATH500. For knowledge-heavy tasks, conformity can perform better than anti-conformity, likely because knowledge gaps make anti-conformity switching more random. The framework is also reported to be insensitive to the number of debate rounds: the \(f_k\) downweighting and the weight vector \(\mathcal{W}\) balance the influence of late-stage changes against early independent reasoning, so single-round debate suffices in the main experimental setting.

The paper also identifies several limitations. A single global weight configuration \(\mathcal{W}\) is used across tasks. Some weaker models under anti-conformity can be overly rigid. Future work is described as including broader agent heterogeneity, larger and harder benchmarks such as HLE, more attack types including prompt injection and adaptive attacks, and systematic exploration of \(\mathcal{W}\) [2509.11035].

## 6. Position within the MAD literature

Within LLM reasoning, Free-MAD occupies a specific position: it is not merely a training-free or efficiency-oriented debate pipeline, but a consensus-free debate framework whose key innovation is a trajectory-aware decision rule. A nearby line of work, ARMOR-MAD, treats debate as conditional computation through Pre-debate Agreement Routing, Early Agreement Stopping Evaluator, and Semantic Outlier Detection. ARMOR-MAD is explicitly training-free and heterogeneous, and it uses agreement-based routing, convergence-based stopping, and outlier-aware aggregation. Free-MAD, by contrast, eliminates the need for consensus among agents, reconstructs the debate phase by introducing anti-conformity, and evaluates the entire debate trajectory rather than relying on the last round only [2606.13197].

That contrast clarifies the scope of the term. In formal usage, Free-MAD denotes the framework titled “Free-MAD: Consensus-Free Multi-Agent Debate,” with two modes, Free-MAD-n and Free-MAD-c, and a score-based decision mechanism over debate trajectories. In a broader interpretive sense, other work can treat “free” as training-free, adaptive, or consensus-light, but Free-MAD’s defining property is the removal of consensus as the organizing principle of debate. The framework therefore replaces majority-driven convergence with a deterministic assessment of how answers are generated, maintained, adopted, or abandoned across the full debate process.

Source: https://www.emergentmind.com/topics/free-mad