MArgE: Multi-LLM Justifiable Claim Verification

• The framework’s main contribution is structuring multi-LLM outputs into formal argument trees for verifiable and justifiable claim verification.
• MArgE employs ArgLLMs based on computational argumentation to extract and organize evidence, ensuring an inspectable pathway from input to decision.
• Empirical results demonstrate that MArgE outperforms single-model approaches and unstructured debates, highlighting the efficacy of formal argumentative reasoning.

Searching arXiv for the target paper and closely related work on argumentative LLMs and multi-LLM claim verification. MArgE, short for “Meshing Argumentative Evidence from Multiple LLMs for Justifiable Claim Verification,” is a framework for claim verification that combines outputs from multiple LLMs while imposing formal argumentative structure on the evidence each model contributes. Its central premise is that multi-LLM aggregation should not be reduced to unstructured interaction such as free debate; instead, each model’s contribution is represented as a tree of extracted arguments, constructed with a variant of Argumentative LLMs (ArgLLMs) grounded in computational argumentation. The resulting process is intended to yield an inspectable pathway from initial arguments to final claim-verification decisions, thereby supporting faithful justification rather than only a final label (Ng et al., 4 Aug 2025).

1. Problem Setting and Motivation

MArgE addresses two linked issues in contemporary LLM-based verification. First, leveraging outputs from multiple LLMs is increasingly used to harness model capability across tasks while mitigating errors such as hallucinations. Second, existing approaches to combining insights from multiple LLMs often rely on unstructured interactions, including free debate, and therefore can produce generations that are not faithfully justifiable (Ng et al., 4 Aug 2025).

Within that framing, claim verification is treated not merely as classification but as a justification-sensitive inference problem. The emphasis is not only on whether a claim is verified, but on whether the route to that decision can be inspected and defended. This places MArgE within a line of work that treats explanation as part of the inferential artifact rather than as an after-the-fact textual gloss. A plausible implication is that the framework is meant to reduce the gap between model output quality and epistemic accountability in verification settings.

2. Core Representational Idea

The defining representational choice in MArgE is to provide formal structure to the evidence from each LLM “in the form of a tree of extracted arguments” (Ng et al., 4 Aug 2025). Rather than letting models exchange unrestricted natural-language responses, the framework extracts and organizes argumentative content into structured trees associated with a given claim.

This structure is the basis for what the title calls “meshing” argumentative evidence from multiple LLMs. The abstract establishes that evidence is collected separately from multiple models and then handled in a formally structured manner. This suggests that inter-model combination is mediated by argument structure rather than by raw conversational interaction. The significance of this design is methodological: it replaces opaque multi-agent interaction with a representation that is intended to be inspectable, compositional, and aligned with argument-based reasoning.

3. ArgLLMs and Computational Argumentation

MArgE uses “a variant of Argumentative LLMs (ArgLLMs), i.e. LLMs driven by frameworks and semantics from the field of computational argumentation,” to construct structured argument trees for claims (Ng et al., 4 Aug 2025). This is the most explicit indication of its theoretical lineage. The framework therefore does not treat argumentation as a stylistic prompt format; it situates the LLM within formal machinery associated with computational argumentation.

That choice matters because the contribution is not simply to ask multiple models for reasons pro and con. Instead, MArgE imports argumentative frameworks and semantics as organizing principles for evidence construction. The paper’s positioning implies that formal argumentation is used to discipline evidence extraction and aggregation, and to connect intermediate argumentative units to a final verification outcome. This suggests a hybrid architecture in which LLMs supply linguistic and inferential content, while computational argumentation supplies structure and evaluative constraints.

4. Justifiable Claim Verification

The target task is claim verification, but the framework’s distinctive claim is that verification should be justifiable in a faithful sense. The abstract states that the process “creates an inspectable pathway from the initial arguments to the final claim verification decisions, providing a faithful justification thereof” (Ng et al., 4 Aug 2025).

In this formulation, justification is not merely an auxiliary explanation generated after the verdict. It is tied to the pathway by which the verdict is reached. For verification systems, this is a substantial distinction. Post hoc explanations may be fluent while failing to reflect the model’s actual decision process; MArgE is presented as an alternative in which the evidential path itself is structured and inspectable. This suggests that the framework is aimed at settings where auditability of reasoning is at least as important as raw predictive performance.

5. Experimental Positioning and Reported Results

The reported empirical claim is that MArgE “can significantly outperform single LLMs,” including “three open-source models (4B to 8B parameters), GPT-4o-mini and existing ArgLLMs,” as well as “prior methods for unstructured multi-LLM debates” (Ng et al., 4 Aug 2025). The comparison set is notable because it spans both single-model baselines and alternative multi-model aggregation strategies.

The paper’s stated conclusion is that these results “demonstrate the advantages of incorporating formal, argumentative reasoning mechanisms when combining multiple LLM outputs” (Ng et al., 4 Aug 2025). Interpreted conservatively, this positions MArgE as evidence that structure matters in multi-LLM systems: gains are attributed not only to model multiplicity, but to the use of formal argumentative reasoning mechanisms. A common misconception in this area is that multi-LLM debate is sufficient by itself to produce reliable or accountable reasoning; MArgE is presented as a direct challenge to that assumption.

6. Conceptual Significance and Scope

MArgE occupies an intersection of multi-LLM orchestration, computational argumentation, and claim verification. Its contribution, as described in the abstract, is to show that combining multiple LLMs need not force a choice between performance and inspectability. By structuring evidence as argument trees and linking those trees to final verification decisions, the framework is explicitly aimed at faithful justification rather than only task accuracy (Ng et al., 4 Aug 2025).

This suggests broader relevance to research on verifiable reasoning, evidence-grounded NLP, and structured decision support. At the same time, the paper’s public summary emphasizes comparative performance and architectural framing rather than detailed formal exposition. Accordingly, the clearest established properties are the use of structured argument trees, the dependence on a variant of ArgLLMs, the production of inspectable pathways, and the reported superiority over both single-model baselines and unstructured debate-based methods.

7. Disambiguation of the Name

The label “MArgE” is ambiguous across the arXiv literature and should be distinguished from several unrelated methods and systems. In particular, MArgE for justifiable claim verification (Ng et al., 4 Aug 2025) is distinct from MaRGE for query-focused summarization (Xu et al., 2020), MARGE for guided exploration in mathematical reasoning (Gao et al., 18 May 2025), MARGE as a multilingual retrieval-and-generation pre-training method (Lewis et al., 2020), MArgE for semi-supervised Gaussian mixture modelling under Missing at Random (Liao et al., 21 Jan 2026), MARgE as a reactive-graphs visualization and analysis tool (Tinoco et al., 2024), and MaRGE as shorthand for the Maxey–Riley–Gatignol equations in particle dynamics (Sommer et al., 9 Apr 2026).

Name	Domain	Paper
MArgE	Justifiable claim verification with multiple LLMs	(Ng et al., 4 Aug 2025)
MaRGE	Query-focused summarization	(Xu et al., 2020)
MARGE	Math reasoning with guided exploration	(Gao et al., 18 May 2025)
MARGE	Multilingual retrieval-and-generation pre-training	(Lewis et al., 2020)
MArgE	Semi-supervised Gaussian mixture modelling under MAR	(Liao et al., 21 Jan 2026)
MARgE	Reactive-graphs tool	(Tinoco et al., 2024)
MaRGE	Maxey–Riley–Gatignol equations	(Sommer et al., 9 Apr 2026)

The claim-verification framework is therefore best identified by its full title, “MArgE: Meshing Argumentative Evidence from Multiple LLMs for Justifiable Claim Verification,” to avoid confusion with these homonymous or near-homonymous usages.