Enhancing Multi-Agent Debate System Performance via Confidence Expression

Zijie Lin, Bryan Hooi

Abstract: Generative Large Language Models (LLMs) have demonstrated remarkable performance across a wide range of tasks. Recent research has introduced Multi-Agent Debate (MAD) systems, which leverage multiple LLMs to simulate human debate and thereby improve task performance. However, while some LLMs may possess superior knowledge or reasoning capabilities for specific tasks, they often struggle to clearly communicate this advantage during debates, in part due to a lack of confidence expression. Moreover, inappropriate confidence expression can cause agents in MAD systems to either stubbornly maintain incorrect beliefs or converge prematurely on suboptimal answers, ultimately reducing debate effectiveness and overall system performance. To address these challenges, we propose incorporating confidence expression into MAD systems to allow LLMs to explicitly communicate their confidence levels. To validate this approach, we develop ConfMAD, a MAD framework that integrates confidence expression throughout the debate process. Experimental results demonstrate the effectiveness of our method, and we further analyze how confidence influences debate dynamics, offering insights into the design of confidence-aware MAD systems.

1. Introduction

Multi-Agent Debate (MAD) systems leverage the collective reasoning of multiple Large Language Models (LLMs) to solve complex tasks. While promising, these systems often suffer from a "stubbornness" problem where agents fail to reach a correct consensus. A primary cause is the lack of a mechanism for agents to communicate their certainty. Without explicit confidence expression, a model with a correct but nuanced answer may fail to convince a peer who is confidently wrong, or the debate may end in a stalemate. This paper introduces ConfMAD, a framework that integrates calibrated confidence scores into the debate process to improve accuracy and consensus.

2. The ConfMAD Framework

ConfMAD enhances the standard debate workflow by requiring agents to express and calibrate their confidence levels. The framework consists of two main stages:

  1. Confidence Expression: Agents provide a confidence score (0–100) alongside their reasoning. The study explores two elicitation methods: LN Confidence, derived from the log-probabilities of generated tokens, and SV Confidence (Self-Verbalized), where the model is directly prompted to state its confidence level.
  2. Calibration Schemes: Since raw LLM confidence is often poorly calibrated (typically overconfident), ConfMAD applies post-hoc calibration. Techniques such as Platt Scaling, Temperature Scaling, and Histogram Binning are used to map raw scores to empirical probabilities, ensuring that the communicated confidence is a reliable indicator of correctness.
The ConfMAD framework workflow, illustrating the initial round of independent opinion generation followed by iterative debate rounds with confidence calibration.

The debate follows a structured "one-by-one" format. In the initial round, agents generate responses independently. In subsequent rounds, agents review the full debate history—including the calibrated confidence scores of their peers—and update their arguments accordingly.

3. Experimental Setup

The framework was evaluated using GPT-4o-mini as the primary agent, paired with either LLaMA-3.1-70B-Instruct or Phi-4. Testing was conducted across four diverse reasoning benchmarks:

  • BIGGSM: Complex mathematical computations.
  • Big-Bench-Hard (BBH): General reasoning tasks.
  • MMLU: Multitask language understanding across 57 subjects.
  • MATH: High-school competition-level mathematics.

4. Results and Analysis

Experimental results demonstrate that incorporating calibrated confidence significantly boosts performance across all benchmarks.

Accuracy comparison across four benchmarks showing the performance of ConfMAD with different calibration schemes against the 'No Confidence Debate' baseline.

Key findings include:

  • Accuracy Gains: ConfMAD consistently outperformed baseline debates that lacked confidence expression. For instance, the SV+Platt configuration achieved a 5.5% accuracy improvement over the baseline on the BIGGSM dataset.
  • The Role of Calibration: Raw self-verbalized confidence was found to be unreliable, as models often claim high certainty even when incorrect. However, applying Platt Scaling effectively "normalized" these scores, making them a useful signal for peer agents.
  • Correction and Consensus: The inclusion of confidence scores acted as a "correction" signal. On the MMLU dataset, the number of cases where an initially incorrect agent was corrected by a peer increased by approximately 20%. Furthermore, ConfMAD reduced the frequency of "agree to disagree" outcomes, leading to an 11% increase in consensus rates.

5. Conclusion

ConfMAD addresses the limitations of traditional MAD systems by enabling agents to communicate their certainty through calibrated confidence scores. This approach not only improves task accuracy but also facilitates more effective debate dynamics by breaking deadlocks and encouraging self-correction. Future work will focus on improving the generalizability of calibration models across different task distributions and exploring dynamic stopping criteria for debates based on confidence stability.

Summary updated

Abstract

Generative Large Language Models (LLMs) have demonstrated remarkable performance across a wide range of tasks. Recent research has introduced Multi-Agent Debate (MAD) systems, which leverage multiple LLMs to simulate human debate and thereby improve task performance. However, while some LLMs may possess superior knowledge or reasoning capabilities for specific tasks, they often struggle to clearly communicate this advantage during debates, in part due to a lack of confidence expression. Moreover, inappropriate confidence expression can cause agents in MAD systems to either stubbornly maintain incorrect beliefs or converge prematurely on suboptimal answers, ultimately reducing debate effectiveness and overall system performance. To address these challenges, we propose incorporating confidence expression into MAD systems to allow LLMs to explicitly communicate their confidence levels. To validate this approach, we develop ConfMAD, a MAD framework that integrates confidence expression throughout the debate process. Experimental results demonstrate the effectiveness of our method, and we further analyze how confidence influences debate dynamics, offering insights into the design of confidence-aware MAD systems.
BibTeX
@inproceedings{lin2025enhancing,
  author = {Zijie Lin and Bryan Hooi},
  title = {{Enhancing Multi-Agent Debate System Performance via Confidence Expression}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure