Abstract
Standard Multi-Agent Debate (MAD) often fails because Large Language Models (LLMs) fall into "belief entrenchment," where they reinforce each other's mistakes rather than correcting them. This paper introduces DReaMAD, a framework that breaks this echo chamber by eliciting strategic prior knowledge and forcing perspective diversity. Evaluated on the MetaNIM Arena benchmark and complex mathematical reasoning tasks, DReaMAD leads to significant performance gains, including a +19.0% higher win rate than standard MAD in strategic games.
Introduction: The Problem of Belief Entrenchment
While Multi-Agent Debate (MAD) was designed to improve LLM reasoning through peer critique, it often suffers from a critical failure mode: belief entrenchment. Instead of filtering for truth, MAD can act as a megaphone for the majority. If agents begin with a shared biased or suboptimal initial thought, the debate process reinforces the error rather than correcting it.
The authors identify two root causes for this phenomenon:
- Static Initial Belief: A model’s inherent bias toward immediate context over long-term strategy.
- Homogenized Debate Dynamics: A "popularity contest" where persuasiveness is driven by consensus rather than logical validity.
The DReaMAD Framework
To counteract entrenchment, the authors propose DReaMAD (Diverse Reasoning via Multi-Agent Debate), a two-stage framework designed to shift the model's reasoning process toward optimality.
- Stage 1: Strategic Prior Knowledge Elicitation (SPKE): Before debating, the model performs "Game Situation Reinterpretation" and "General Strategy Formulation." By articulating rules and winning conditions first, the model shifts its initial probability distribution toward the correct action.
- Stage 2: Perspective Diversification: Each agent is assigned a unique, self-generated prompt to encourage exploration of different reasoning paths, ensuring consensus is built on logic rather than repetition.
The framework is grounded in the Sprague-Grundy Theorem, allowing for objective measurement of move optimality in impartial games.
Experimental Setup: MetaNIM Arena
The researchers introduced MetaNIM Arena, a rigorous benchmark consisting of five adversarial strategic games: NIM, Fibonacci Nim, Kayles, Chomp, and Corner Queen. These games test various aspects of reasoning, from state dependency to problem decomposition. Additionally, the framework was tested on high-level math benchmarks (AIME 2024, AMC 2023) and CommonsenseQA.
Results and Discussion
DReaMAD significantly outperforms standard prompting and debate methods. It achieved a +9.5% accuracy gain over ReAct prompting on MetaNIM Arena and a +19.0% higher win rate than standard MAD against a strong GPT-4o opponent.
Key insights include:
- The Popularity Trap: Mathematical modeling shows that the probability of accepting an argument is often proportional to the initial belief, biasing standard MAD toward "first impressions."
- Diversity as Error-Correction: By forcing agents to adopt different perspectives, the debate transforms into a robust error-correction mechanism.
- Unlocking "Thinking" Capabilities: On the Qwen3-4B model, DReaMAD provided a 45.1% performance gain, allowing a small base model to rival specialized "Thinking" models without additional training.
Limitations and Future Work
DReaMAD's effectiveness is tied to the availability of a strategic foundation; it struggled with games like Chomp that lack simple closed-form solutions. Future research may explore heterogeneous ensembles (mixing different models) to further mitigate entrenchment. While highly effective, the framework does involve increased computational overhead due to multiple interaction steps.
Summary updated