Large Language Models (LLMs) frequently produce fluent but unverifiable reasoning, resulting in potential hallucinations and faulty inferences. This study proposes an argumentation-based verification framework ValidArgLLM in which the reasoning expressed by an LLM is transformed into a defeasible logic program (DLP) representing world knowledge and a given problem description-such as a patient health complaint. The DLP is executed within a symbolic reasoning engine, and the resulting inferences are compared to the LLM's natural-language conclusions. The strength of arguments is computed based on discourse structure of text expressing arguments. Divergence between symbolic and neural reasoning outcomes indicates possible hallucination or inconsistency in the model's internal logic.
Large language models (LLMs) have achieved impressive results across diverse natural language
processing tasks, inspiring interest in their use for domains that require structured reasoning.
However, integrating LLMs into settings that demand context-sensitive, multi-step decision-making
remains challenging. These models often excel at generating fluent and informed text but struggle
to reason systematically, weigh competing possibilities, or revise conclusions when new information
appears [
Another major limitation is interpretability. Unlike human experts who reason through explicit, traceable arguments, LLMs reach conclusions through opaque statistical processes [22]. This opacity makes it difficult to understand or justify their outputs, reducing trust and accountability in applications that require verifiable logic [23]. Furthermore, LLMs are prone to reasoning hallucinations—producing statements that sound coherent but conflict with facts or internal logic. Because they lack explicit mechanisms for defeasibility or conflict resolution [18, 19, 20, 21], such inconsistencies can undermine reliability. To address these gaps, LLMs must be complemented by external reasoning and verification layers capable of enforcing logical consistency and explaining why conclusions hold.
One promising approach is to pair an LLM with a symbolic or logical reasoning engine—for
example, a Prolog-style rule base, a constraint solver, or a formal ontology of medical conditions [
To mitigate these risks, we introduce a neuro-symbolic verification framework ValidArgLLM that externalizes and tests an LLM’s reasoning through argumentation analysis. The key idea is to translate the model’s implicit causal and conditional reasoning into a formal system—defeasible logic programming (DeLP)—and verify its conclusions via a logical solver.
Despite significant advancements in task performance, current approaches to argumentative
reasoning continue to face critical limitations in terms of justifiability and interpretability. While
models can often produce seemingly coherent explanations, it remains ambiguous how final
decisions are actually reached or which intermediate steps genuinely contributed to the outcome.
Recent studies have revealed that Chain-of-Thought (CoT) reasoning, though designed to enhance
transparency, is itself prone to hallucinations, inconsistencies, and spurious reasoning paths that do
not accurately represent the underlying inference mechanisms of large language models [
Figure 1 compares four levels of reasoning used by language models, showing how they evolve
from simple answers to structured, self-correcting discourse-based reasoning [
The next level, chain-of-thought reasoning, adds a sequence of intermediate steps that make the process more interpretable. However, these steps remain unverified. The reasoning might sound plausible while still being factually wrong, since the model does not test or challenge its own statements.
The argumentative model introduces a more structured approach. Instead of producing one line of reasoning, it generates multiple arguments, distinguishing between supporting and attacking ones (Gutiérrez et al. 2024). This enables a form of contestability: each conclusion is backed by explicit evidence and can be challenged by counterarguments. Still, this stage only formalizes reasoning—it does not validate or improve it. The model outputs argument structures but lacks a mechanism to revise its own conclusions.
knee pain. subsided. hTehaertpaattiteanckt.experienced a iSmnpyfeeocrciifoairrcdawilalayll,li.nafnaorcnt-iSoTn aeflefevcattiinognthe The patient’s cholesterol Although his diet compliance was level improved. inconsistent.
The surgeon performed an The patient had arrived at the hospital emergency appendectomy. only two hours earlier.
The nurse administered a To facilitate the insertion of a central sedative. venous catheter.
The patient is diagnosed Supported by chest X-ray showing with pneumonia. bilateral infiltrates.
The treatment outcome is According to the hospital’s quality considered successful. benchmarks.
The discourse-based argument validation model ValidArgLLM at the bottom represents the most advanced form. It integrates discourse structure to evaluate how strongly each argument contributes to the overall reasoning. By analyzing rhetorical relations such as elaboration, justification, or background, it can weigh the importance of each argument and decide which should dominate the conclusion. This allows the system to detect when the original model’s answer is inconsistent with the discourse-level balance of evidence and automatically correct it. Over time, this validation loop enables the model to refine its reasoning and become more consistent and interpretable. This architecture adds discourse awareness, verifiable structure, and self-correction. It moves beyond merely producing or scoring arguments—it understands how arguments interact and uses that understanding to ensure that reasoning outcomes are both coherent and defensible.
Each rhetorical relation has a nucleus (the “main” proposition) and a satellite (supporting or contextual material). The satellite always carries less essential information than the nucleus (Table 1). One can see that nucleus contains main diagnostic/treatment fact (higher base probability) and satellite caries contextual/supporting info with lower significance. These values are obtained in the course of improvement of validation performance, described in Evaluation section.
1) strict rules (must hold, nucleus) Head <- Body. 2) defeasible rules (should hold, satellite) Head <~ Body.
fact. % strict fact vs fact <~ . % defeasible fact We now show how a defeasible logic program in ValidArgLLM would be built from a real nucleus vs. satellite pair.
Nucleus (main claim): “The patient must immediately start a course of antibiotics for bacterial pneumonia.” Satellite (supporting context): “Because the chest X-ray shows an infiltrate consistent with pneumonia.”
Here the nucleus expresses a mandatory action (antibiotics). The satellite is evidence/justification (X-ray finding) — useful, but not itself the main prescription.
In a defeasible logic program: % Strict rule from the nucleus: MUST do this if bacterial pneumonia diagnosed start_antibiotics(Patient) <- diagnosis(Patient, bacterial_pneumonia). % Defeasible rule from the satellite: SHOULD suspect pneumonia if X-ray shows infiltrate diagnosis(Patient, bacterial_pneumonia) <~ chest_xray_infiltrate(Patient). % Facts: chest_xray_infiltrate(john) <~ . % defeasible evidence The nucleus states the obligatory/primary outcome, while the satellite is only supportive, so its rule is defeasible/overridable.
The nucleus carries the main point of the discourse. In DLP you can treat it as a strict rule or a strict fact, because it is asserted to hold independently of the supporting material. diagnosis(Patient, bacterial_pneumonia).
Even if we later remove the satellite (the X-ray), the nucleus stands on its own.
The satellite carries supporting or contextual information. In DLP you model it as a defeasible rule whose conclusion only fires under the context of the nucleus (or some nucleus-derived condition).
% satellite only applies if nucleus (diagnosis) is already true chest_xray_infiltrate(Patient) <~ diagnosis(Patient, bacterial_pneumonia).
More generally: satellite_fact(X) <~ nucleus_fact(X).
So, in ValidArgLLM the satellite is not automatically accepted; it’s accepted defeasibly and only when its nucleus context holds. In a defeasible logic program derived from discourse, nucleus statements are converted into strict rules or strict facts that hold independently. Satellite statements are converted into defeasible rules whose applicability is conditional on the corresponding nucleus being true; they express “should” or “likely” information that can be overridden or becomes vacuous if the nucleus is absent.
Given a natural-language case description, the LLM within ValidArgLLM constructs a set of defeasible rules encoding its inferred world knowledge: r1: gout :- asymmetric_joint_inflammation, uric_acid_high. r2: immune_arthritis :- symmetric_joint_inflammation, fever. r3: asymmetric_joint_inflammation :- not symmetric_joint_inflammation. r4: prefer immune_arthritis over gout if fever.
Here, each rule expresses a conditional belief that may be overridden by stronger evidence. A defeasible logic program is a set of facts, strict rules, Π, of the form A:-B, and a set of defeasible rules Δ of the form, A- < B, whose intended meaning is “if B is the case, then usually A is also the case.” A DeLP for knowledge sources includes facts which are extracted from search results and strict and defeasible clauses where the head and body form commonsense reasoning rules (Garcia and Simari, 2004).
Let DT=(N,R) be a discourse tree, where N is the set of elementary discourse units (EDUs), and R⊆N×N is the set of rhetorical relations between nucleus and satellite spans. Each relation !=(nucleus,satellite,relation) ∈R has a rhetorical type (e.g., Cause, Evidence, Elaboration, Concession) and an associated relative argument strength coefficient ∝"#$%&!'( ∈ , representing how much more influential the nucleus is compared to its satellite.
For every defeasible rule A -< B1,B2,…,Bk ∈ Δ we associate a discourse strength weight w(A)∈ computed as w(A) = |+%&)(-)| ∑((/0 ,2%&,"#$%&!'( ∈4! ∝"#$%&!'( , where - is the set of discourse relations in which participates, and Sat(A) are its supporting satellites. Thus, rules derived from nucleus EDUs obtain higher w(A) values (closer to 1), while rules from satellite EDUs obtain lower w(A)proportionally to their rhetorical importance. Let P=(Π, Δ) be a DeLP program and L a ground literal. A defeasible derivation of L from P consists of a finite sequence L1, L2,…, Ln of ground literals, such that each literal Li is in the sequence because:
2. there exists a rule Ri in P (strict or defeasible) with head Li and body B1, B2,…,Bk and every literal of the body is an element Li of the sequence appearing before Lj (j<i).
Let h be a literal, and P=(Π, Δ) a DeLP program. We say that <A, h> is an argument for h, if A is a set of defeasible rules of Δ, such that: Hence an argument <A, h> is a minimal noncontradictory set of defeasible rules, obtained from a defeasible derivation for a given literal h associated with a program P.
The generated DLP is executed within a defeasible reasoning environment which is the SWIProlog extension. Each rule is treated as an argument, and conflicts among arguments give rise to a dialectical structure that determines which conclusions are warranted under a chosen semantics (e.g., grounded or preferred extension).
The logical conclusions derived by the solver are compared with the LLM’s original verbal output. If both converge (e.g., both conclude immune arthritis), the LLM’s reasoning is considered grounded. If they diverge (e.g., solver: immune arthritis, LLM: gout), the discrepancy signals a reasoning hallucination—a claim unsupported by the formal reconstruction of its own reasoning.
We represent a quantitative bipolar argumentation framework (QBAF [
Gradual semantics recursively compute an argument’s dialectical strength by combining its base
score with the aggregated strengths of its attackers and supporters. Given a gradual semantics, such
as DF-QuAD [
Q(x)≥0.5 or false otherwise.
The strength aggregation function is defined as : *→ , where for a permutation of arguments S=(v1, . . . , vn) ∈ *: if n = 0 ∶ (S) = 0 ; if n = 1 ∶ (S) = v1; if n = 2 ∶ (S) = f (v1,v2) if n > 2 ∶ (S) = f (s(v1, . . . ,vn-1),vn) with the base function f ∶ × → defined, for v1,v2 ∈ , as: f(v1,v2) = v1+ (1−v1)⋅v2 = v1+v1 − v1⋅v2.
Thus, the base function expresses sequences of strengths of attackers or supporters by s,
proportionally increasing the attacking or supporting arguments’ strength towards 1 [
Fig. 2 illustrates a hybrid neuro-symbolic diagnostic reasoning pipeline where an LLM and a Defeasible Logic Programming (DeLP) reasoning engine work together to verify or refute an LLMgenerated medical diagnosis. The goal of the architecture is to ensure that an LLM’s diagnostic answer (for example, “The patient has gout”) is not only linguistically plausible but also logically justified and consistent with a structured medical ontology. If the logical reasoning pipeline cannot defeat the diagnosis claim, it is confirmed; otherwise, the LLM answer is marked unconfirmed. The ValidArgLLM‘s workflow is as follows: 1. User Input. The process begins with a user request, such as asking the model to provide a diagnosis. 2. LLM Generates Initial Answer. The LLM processes the request and outputs an initial diagnosis or conclusion (e.g., “The disease is gout”). 3. Ontology and Discourse Setup. A textual ontology of medical knowledge (rules, relationships, symptoms, conditions) and a discourse parser are available. The discourse parser identifies rhetorical relations in the text — for instance, nucleus (main facts) and satellite (contextual or defeasible facts). Nucleus → “Must” clauses (non-defeasible rules) and Satellite → “Should” clauses (defeasible rules) 4. LLM forms ontology representation, transforming textual information into a rule-based ontology in the DeLP format — essentially translating natural-language reasoning into structured logical rules. 5. Conversion to defeasible logic. The LLM converts these rules into regular (strict) or defeasible (soft) clauses depending on their role in the discourse (main vs. secondary information). 6. Building logical representation of the user request. The system formalizes the user’s question and the LLM’s proposed answer as a set of logical facts that can either be defeated or supported by the ontology. 7. Discourse representation integration. The LLM builds discourse representations of both the user request and the ontology, capturing which arguments are more or less important (nucleus/satellite weighting) and how they relate contextually. 8. Argumentation Pipeline. The argumentation module computes attack relations among rules (contradictions or counter-arguments), dialectical trees, representing possible argumentative dialogues between supporting and opposing claims, and defeasibility outcomes, determining whether a claim survives all counter-arguments 9. Comparison and validation. The LLM compares its original diagnosis with the verified diagnosis obtained through the logical argumentation process. 10. Decision. If the logical reasoning shows that the diagnosis claim is not defeated, it is confirmed as valid. If the claim is defeated by stronger counter-arguments from the ontology, it is marked unconfirmed.
• • •
Neural generation (LLM) → producing hypotheses and natural-language reasoning. Symbolic verification (DeLP) → testing those hypotheses for logical soundness.
Discourse analysis → weighting arguments by rhetorical importance.
Together, these components produce a contestable and interpretable diagnostic system, where the model’s answer can be justified or overturned through structured reasoning.
The demo of argument-based LLM verifier is available at Tool Series for LLM Verification (Figure 3 and Figure 4).
We evaluate on three claim-verification datasets that we derive from existing QA/NLI resources:
TruthfulHalluc (from TruthfulQA [
Our focus is hallucination detection for model answers using argumentation analysis as a validator. The validator assesses whether an answer’s central claim is defeated by the argumentvalidation system. We define a hallucination as a claim whose defeat probability exceeds 0.5. This cautious threshold is motivated by safety-critical domains (health, legal, finance), where we prefer to reject answers that are defeated with substantial probability.
Dataset size and prevalence are as follows. Each hallucination dataset contains 1,000 QA pairs with a 50% hallucination rate (balanced positives/negatives). In the original source datasets the natural hallucination rate is <1%; our perturbation procedure raises prevalence to enable meaningful detection metrics and comparability with prior LLM-argumentation studies.
We report F1 for hallucination prediction in Table 2. It lists: (i) a GPT-5 baseline; (ii) our claimverification tool that uses argument validation; and (iii) a discourse-aware variant where argument strength additionally incorporates discourse cues beyond the default computation. This final column shows the incremental gains from discourse-informed argument strength.
Our MedHalluc results are broadly comparable to prior work: ArgMed-Agents with GPT-4 reports
0.91 predictive accuracy [16]; ArgLLM with GPT-4o reports 0.80 (Friedman et al., 2015); and an
ensemble of ArgLLMs achieves 0.73 [
The approach of Bezou-Vrakatseli (2023) [
This verification strategy directly supports the project’s broader goal of fostering ethical debate between humans and AI systems. By evaluating LLMs through the lens of argumentation theory, researchers can determine how well these models engage in principled reasoning about moral dilemmas, identify potential biases or logical fallacies, and measure their capacity to both construct and critique ethical arguments. Ultimately, this contributes to “ethics for AI”—ensuring AI systems behave responsibly—and “AI for ethics”—using AI as a tool to help humans reflect on and refine their own ethical reasoning. Such a dual focus positions LLMs not just as information providers, but as collaborative partners in navigating complex moral questions.
Earlier approaches to argumentative reasoning, such as ArgLLMs [
These techniques diversified the evaluation signal and reduced model-specific bias but remained primarily semantic—focused on the content of individual arguments rather than their rhetorical role in a discourse structure. In contrast, our work introduces a discourse-based scoring framework that evaluates arguments in the context of their rhetorical relations and structural significance within the discourse tree. Instead of relying solely on model-elicited or reward-style judgments, scores are inferred from the hierarchical organization of argumentative elements—linking claims, evidence, and counterarguments through nucleus–satellite dependencies and coherence relations. This approach enables the system to weight contributions based on discourse salience rather than raw textual confidence, thereby capturing how strongly each component supports or undermines the overall claim.
While reward-model scoring focuses on factual adequacy and semantic quality, our discoursebased approach emphasizes relational justifiability, integrating structural reasoning to yield a more explainable and linguistically grounded measure of argument strength.
Argument analysis tool belongs to the series of LLM verification tools including logic programming, answer set programming, and rule attenuation (Fig. 3) All these tools rely on discourse analysis to determine rule structure and weights to build a logic program representation. We observed that LLMs can be reliably verified by discourse-based argumentation analysis.
Our evaluation demonstrates that integrating argument-validation into LLM pipelines substantially improves hallucination detection across three newly constructed claim-verification datasets: TruthfulHalluc, MedHalluc, and eSNLI_Halluc. By transforming QA/NLI sources into structured QA pairs and injecting semantically incompatible attributes, we create balanced datasets where hallucinations correspond to claims defeated by an argumentation engine. Using a defeatprobability threshold of 0.5—chosen for safety-critical settings—the argument-validation module yields consistent gains over a GPT-5 baseline. Across datasets, ValidArgLLM increases F1 by +0.15– 0.20, with the largest improvements observed in medically grounded reasoning tasks where unsupported causal links and symptom inferences are more common. These results highlight that logical defeat, rather than surface-level confidence, provides a more robust criterion for identifying model errors in multi-step explanatory contexts.
Adding discourse-aware argument strength further strengthens performance for domains where rhetorical centrality matters. Incorporating nucleus–satellite weighting and discourse-relation cues yields additional gains of +0.03–0.06 F1, reaching 0.78 on TruthfulHalluc and 0.80 on MedHalluc. While superficially comparable to prior systems such as ArgMed-Agents (0.91), these approaches estimate truthfulness, whereas our model predicts hallucination by determining whether the answer’s core claim is logically defeated. The distinction is crucial: truth-prediction assumes access to ground-truth facts, whereas defeat-based hallucination detection evaluates internal argumentative consistency. Our results thus establish discourse-augmented argumentation analysis as an effective, model-agnostic verification layer for improving LLM reliability in high-stakes reasoning environments.
During the preparation of this work, the author used GPT 5 in order to: grammar and spelling check, paraphrase and reword. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content. [15] Jin D, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. What disease does this patient have? A large-scale open domain question answering dataset from medical exams. CoRR, abs/2009.13081, 2020. [16] Hong S, Liang Xiao, Xin Zhang, Jianxia Chen (2024) ArgMed-Agents: Explainable Clinical
Decision Reasoning with LLM Disscusion via Argumentation Schemes [17] Camburu O-M, Tim Rocktäschel, Thomas Lukasiewicz, Phil Blunsom (2018) e-SNLI: Natural Language Inference with Natural Language Explanations. Advances in Neural Information Processing Systems 31 (NeurIPS 2018) [18] Ruiz-Dolz R and Lawrence J. 2023. Detecting Argumentative Fallacies in the Wild: Problems and Limitations of Large Language Models. In Proceedings of the 10th Workshop on Argument Mining, pages 1–10, Singapore. Association for Computational Linguistics [19] Xu Z, Sanjay Jain, Mohan Kankanhalli (2024) Hallucination is Inevitable: An Innate Limitation of Large Language Models. arXiv:2401.11817 [20] Banerjee S, Ayushi Agarwal, Saloni Singla (2024) LLMs Will Always Hallucinate, and We Need to Live With This. arXiv:2409.05746 [21] Gutiérrez A, Stella Heras and Javier Palanca (2024) Detecting disinformation through computational argumentation techniques and large language models. CMNA 2024 [22] Musi E and Rudi Palmieri (2024) The Fallacy of Explainable Generative AI: evidence from argumentative prompting in two domains. CMNA 2024 [23] Prakken H (2024) On Evaluating Legal-Reasoning Capabilities of Generative AI. CMNA 2024