From Belief Entrenchment to Robust Reasoning in LLM Agents

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.

💡 Research Summary

The paper investigates a critical failure mode of Multi‑Agent Debate (MAD) for large language models (LLMs), which the authors term “belief entrenchment.” In this phenomenon, agents do not correct each other’s mistakes; instead, they amplify a shared bias, leading to sub‑optimal consensus, especially in sequential, strategic tasks. The authors decompose the problem into two root causes: (1) a biased static initial belief (the probability distribution p(0) over possible actions that the model forms before any interaction) and (2) homogenized debate dynamics, where the persuasiveness of an argument is driven by its initial popularity rather than logical validity.

To formalize the dynamics, they propose a simple probabilistic model of pairwise debate. The probability of action i at round t + 1 is given by p(t+1)ᵢ = p(t)ᵢ·p(t)ᵢ + 2·p(t)ᵢ·(1 − p(t)ᵢ)·p(0)ᵢ. The first term captures agreement, the second term models conflict resolution biased toward the initially popular argument. Empirical measurements on Fibonacci Nim (where the optimal move is mathematically deterministic) show that actions with strong initial consistency (p(0) > 0.5) consistently increase their probability after debate, regardless of correctness. This confirms that standard MAD acts as an “echo chamber” amplifier of the majority view.

To counteract both causes, the authors introduce DReaMAD (Diverse Reasoning via Multi‑Agent Debate with Refined Prompt). It consists of two sequential stages. Stage 1, Strategic Prior Knowledge Elicitation (SPKE), forces the model to reinterpret the game situation and formulate high‑level winning strategies before any debate. By using a low temperature (0.1) and explicit prompts for “Game Situation Reinterpretation” and “General Strategy Formulation,” SPKE shifts the initial belief mass p(0) toward the optimal action. Stage 2, Perspective Diversification, assigns each debating agent a unique, self‑generated prompt, encouraging distinct reasoning paths. A higher temperature (0.7) promotes diversity, breaking the dependence of persuasiveness on p(0) and ensuring that arguments win on logical merit.

The authors evaluate DReaMAD on a newly created benchmark, MetaNIM Arena, which comprises five impartial games (NIM, Fibonacci Nim, Kayles, Chomp, Corner Queen). The benchmark provides both a static dataset of critical game states (for decision‑accuracy measurement) and an interactive simulator where agents face a strong GPT‑4o opponent using the ReAct framework. Results show that DReaMAD improves optimal‑action selection accuracy by +9.5 % over standard ReAct prompting and achieves a +19 % higher win rate than vanilla MAD in the simulator. Across multiple LLMs (GPT‑4o, Gemini‑1.5‑pro, Gemini‑1.5‑flash, Qwen‑3‑4B) the SPKE module consistently raises accuracy, while the diversification module prevents convergence to biased consensus. Ablation studies confirm that both stages are necessary: removing SPKE leaves the initial belief biased, and removing diversification restores the echo‑chamber effect.

Additional experiments demonstrate that DReaMAD can generate long chain‑of‑thought reasoning without further fine‑tuning and shows modest transfer to general NLP tasks such as fact verification, though its strongest gains appear in domains with clear optimal solutions.

In summary, the paper provides a rigorous analysis of why standard multi‑agent debate can fail, proposes a theoretically grounded and empirically validated solution, and introduces a challenging new benchmark for strategic reasoning. DReaMAD’s combination of prior‑knowledge elicitation and enforced perspective diversity offers a practical recipe for building more robust, truth‑seeking LLM debate systems, advancing the reliability of AI‑driven decision making in complex, adversarial environments.