Abstract
Multi-Agent Debate (MAD) has emerged as a promising inference scaling method for Large Language Model (LLM) reasoning. However, it frequently suffers from belief entrenchment, where agents reinforce shared errors rather than correcting them. Going beyond merely identifying this failure, we decompose it into two distinct root causes: (1) the model's biased $\textit{static initial belief}$ and (2) $\textit{homogenized debate dynamics}$ that amplify the majority view regardless of correctness. To address these sequentially, we propose $\textbf{DReaMAD}$ ($\textbf{D}$iverse $\textbf{Rea}$soning via $\textbf{M}$ulti-$\textbf{A}$gent $\textbf{D}$ebate with Refined Prompt). Our framework first rectifies the static belief via strategic prior knowledge elicitation, then reshapes the debate dynamics by enforcing perspective diversity. Validated on our new $\textit{MetaNIM Arena}$ benchmark, $\textbf{DReaMAD}$ significantly mitigates entrenchment, achieving a +9.5% accuracy gain over ReAct prompting and a +19.0% higher win rate than standard MAD.
Introduction
While Large Language Models (LLMs) have shown remarkable capabilities, their ability to self-correct during inference remains limited. Multi-Agent Debate (MAD) was introduced to address this by allowing multiple agents to critique and refine each other's reasoning. However, this paper identifies a critical failure mode termed belief entrenchment. In complex strategic tasks, if a majority of agents begin with an incorrect initial belief, the debate process often serves to amplify this error rather than resolve it. We attribute this to the "echo chamber" effect created by homogenized debate dynamics among identical model instances, where persuasiveness is driven by initial popularity rather than logical validity.
The DReaMAD Framework
To overcome belief entrenchment, we propose DReaMAD (Diverse Reasoning via Multi-Agent Debate). The framework introduces two key interventions to break the cycle of error reinforcement:
- Strategic Prior Knowledge Elicitation (SPKE): Before the debate, the model is prompted to "step back" to reinterpret problem constraints and formulate a high-level winning strategy. This triggers latent domain-specific logic—such as the Sprague-Grundy Theorem in strategic games—that might otherwise remain dormant.
- Perspective Diversification: Instead of using identical prompts, the system generates unique, refined prompts for each agent. This forces the ensemble to explore a wider variety of reasoning paths, ensuring that the subsequent debate is grounded in diverse viewpoints rather than a single shared bias.
Experimental Evaluation
We evaluate DReaMAD primarily on the MetaNIM Arena, a new benchmark consisting of five impartial games: NIM, Fibonacci Nim, Kayles, Chomp, and Corner Queen. These games are ideal for testing reasoning because they possess mathematically provable optimal moves, yet present significant challenges for standard LLM prompting. To test broader generalization, we also include mathematics competitions (AIME 2024, AMC 2023) and the CommonsenseQA dataset.
Results and Discussion
Our results demonstrate that DReaMAD significantly outperforms existing methods. It achieves a +9.5% accuracy gain over ReAct prompting and a +19.0% higher win rate than standard MAD.
A key finding is the mitigation of belief entrenchment. As shown in our analysis of decision rates, standard MAD often sees a rapid decline in optimal actions as agents converge on a popular but incorrect "biased" action. In contrast, DReaMAD maintains a high rate of optimal decision-making across multiple rounds of debate by preventing the initial "popularity trap."
Notably, on the Qwen3-1.7B model, DReaMAD provided a 226% average performance gain, elevating the base model's performance to nearly match specialized, heavily post-trained "Thinking" versions. From a cost perspective, DReaMAD is highly efficient, outperforming fine-tuned models at a fraction of the API cost (e.g., $0.0098 vs $0.043 for GPT-4o-mini).
Conclusion
DReaMAD provides a robust framework for unlocking latent reasoning in LLMs without the need for expensive fine-tuning. While the method excels in games with closed-form strategies, challenges remain in heuristic-heavy tasks like Chomp. Future work will explore heterogeneous agent ensembles to further mitigate entrenchment through architectural diversity.
Summary updated