[Paper Review] Examining Inter-Consistency of Large Language Models Collaboration: An In-depth Analysis via Debate

Kai Xiong, Xiao Ding+3 more|arXiv (Cornell University)|May 19, 2023

Topic Modeling4 citations

TL;DR

This paper introduces FORD, a formal debate framework to study inter-consistency in collaborative large language models (LLMs) across commonsense reasoning tasks. Through three-stage debates—fair, mismatched, and roundtable—LLMs achieve consensus despite inter-inconsistencies, though performance is hindered when abilities are imbalanced; using GPT-4 as a judge slightly improves outcomes, highlighting the role of authoritative oversight in collaborative LLM systems.

ABSTRACT

Large Language Models (LLMs) have shown impressive capabilities in various applications, but they still face various inconsistency issues. Existing works primarily focus on the inconsistency issues within a single LLM, while we complementarily explore the inter-consistency among multiple LLMs for collaboration. To examine whether LLMs can collaborate effectively to achieve a consensus for a shared goal, we focus on commonsense reasoning, and introduce a formal debate framework (FORD) to conduct a three-stage debate among LLMs with real-world scenarios alignment: fair debate, mismatched debate, and roundtable debate. Through extensive experiments on various datasets, LLMs can effectively collaborate to reach a consensus despite noticeable inter-inconsistencies, but imbalances in their abilities can lead to domination by superior LLMs. Leveraging a more advanced LLM like GPT-4 as an authoritative judge can boost collaboration performance. Our work contributes to understanding the inter-consistency among LLMs and lays the foundation for developing future collaboration methods. Codes and data are available at https://github.com/Waste-Wood/FORD

Motivation & Objective

• To investigate inter-consistency among multiple LLMs during collaborative reasoning, complementing existing work on single-model self-consistency.
• To design a structured debate framework (FORD) that enables systematic, quantitative analysis of LLM collaboration in real-world aligned scenarios.
• To evaluate how LLMs with varying capabilities interact during debate and whether they can reach a shared, reliable conclusion.
• To assess the impact of a more advanced LLM (GPT-4) as a judge in summarizing debates and improving final consensus.

Proposed method

• Proposes a three-stage debate framework: fair debate (comparable LLMs), mismatched debate (differing LLM capabilities), and roundtable debate (multiple LLMs).
• Uses multi-choice commonsense reasoning datasets (e.g., CSQA, PIQA, StrategyQA) to quantify inter-inconsistency and measure collaboration outcomes.
• Employs a formalized debate protocol where LLMs alternate presenting arguments, counterarguments, and potential compromises.
• Introduces a metric INCON to quantitatively measure inter-inconsistency across LLMs’ predictions.
• Uses GPT-4 as a judge to summarize debates and assign final conclusions, improving consensus reliability.
• Applies a voting mechanism when no agreement is reached, with final answers derived from majority consensus.

Figure 1: (a) compromise and (b) refutation in LLMs debates. is the proponent, and is the opponent.

Experimental results

Research questions

• RQ1 To what extent can multiple LLMs achieve consensus through structured debate, despite inherent inter-inconsistencies?
• RQ2 How do capability imbalances between LLMs affect debate dynamics and final consensus quality?
• RQ3 Can a more advanced LLM (e.g., GPT-4) serve as an effective judge to improve collaboration outcomes?
• RQ4 How do different LLM architectures (e.g., chat vs. text completion) compare in collaborative reasoning under debate?
• RQ5 Does the order of debate or the number of participants significantly affect the final outcome in roundtable debates?

Key findings

• Different LLM types, even from the same base model, exhibit high inter-inconsistency, indicating inherent variability in reasoning behavior.
• Even within the same LLM series (e.g., GPT-3.5), newer versions do not fully replace older ones, as earlier models retain some reasoning capabilities.
• LLMs with comparable abilities can effectively collaborate and reach consensus, achieving high performance on commonsense reasoning tasks.
• In mismatched debates, superior LLMs dominate and resist changing views, while weaker models are more likely to compromise, potentially lowering overall performance.
• Using GPT-4 as a judge improves collaboration performance slightly, with average accuracy rising from 80.05% to 80.23% across datasets.
• Roundtable debates show dominance patterns, with certain LLMs consistently influencing outcomes, suggesting structural biases in multi-agent collaboration.

Figure 2: The overall workflow of FORD on two LLMs. Step 1 : LLMs independently give a choice and explanation for each sample as the stance and argument, respectively. Step 2 : LLMs debate alternately on the inconsistent samples. Step 3 : A judge summarizes the whole debate process and gives the fin