Argumentative LLMs Overview

• Argumentative LLMs are large language models that structure reasoning into explicit argument graphs with supports and attacks, enabling transparent decision-making.
• They integrate formal argumentation frameworks—including Dung-style and quantitative bipolar models—to compute claim acceptability with deterministic or graded semantics.
• Their applications span claim verification, legal reasoning, medicine, and discourse analysis, offering inspectable and contestable outcomes while addressing input sensitivity and computational overhead.

Searching arXiv for papers on Argumentative LLMs and the cited frameworks. Argumentative LLMs (ArgLLMs) are LLM systems in which reasoning is externalized into explicit argumentative structures—arguments, supports, attacks, and, in some variants, quantitative strengths—so that decisions are computed by formal semantics rather than being left as opaque next-token generation. Across recent work, the term covers several related design patterns: formal computational-argumentation plugins that guide answer selection without retraining, multi-agent systems that debate and verify pragmatic classifications, quantitative bipolar argumentation frameworks for explainable decision support, and contestable reasoning interfaces in which users can edit the underlying graph and force recomputation of the outcome (Freedman et al., 2024). In this literature, ArgLLMs are positioned as an attempt to move beyond unstructured chain-of-thought or debate transcripts toward reasoning artifacts that are inspectable, reproducible, and, in some settings, revisable by human users (Vasileiou et al., 16 Mar 2026).

1. Definition, scope, and conceptual boundaries

Argumentative LLMs are defined in closely related but not identical ways across the literature. In claim verification and decision-support settings, they are systems that use LLMs to generate supporting and attacking arguments, assign intrinsic strengths, and then evaluate the resulting structure with formal argumentation semantics to derive a final score or label (Freedman et al., 2024). In legal reasoning, the same paradigm is described as one in which LLMs do not merely produce explanations about their answers but actively construct, organize, and compute over explicit arguments—claims, supports, and attacks—so that the reasoning can be verified, weighed, and contested (Cao et al., 21 Feb 2026). In discourse analysis, the notion is specialized further: ArgLLMs are LLM agents explicitly grounded in argumentation and rephrase theory to analyze rhetorical and pragmatic functions of discourse rather than surface similarity alone (Uberna et al., 14 Feb 2026).

This shared core distinguishes ArgLLMs from several adjacent paradigms. Standard LLM generation typically follows a single answer path and lacks an explicit mechanism for resolving contradictions among candidate claims. Chain-of-thought, self-consistency, reflection, and related prompting methods can improve intermediate reasoning traces, but they do not encode attack and defense relations or compute acceptability under formal semantics (Castagna et al., 2024). Conventional retrieval-augmented generation can supply external evidence, but it does not by itself determine which conflicting claims are defensible according to an argumentation-theoretic rule system. Debate-style multi-agent systems similarly differ when they retain only transcripts rather than a structured argument graph evaluated by an external solver (Jin et al., 29 Jan 2026).

A recurring motivation is the distinction between explanation and faithfulness. Several papers argue that free-form rationales may be persuasive without being causally tied to the final answer, whereas ArgLLMs derive decisions from a structure that is itself displayed to the user (Freedman et al., 2024). This suggests a more restrictive but more auditable notion of reasoning: the system must first externalize arguments and relations, then let a deterministic or quantitatively specified semantics decide what follows.

2. Formal substrates and reasoning semantics

The formal backbone of ArgLLMs is computational argumentation. A standard starting point is Dung-style abstract argumentation, where an argumentation framework is written as $\mathrm{AF} = (A, R)$ with arguments $A$ and attack relation $R \subseteq A \times A$ (Castagna et al., 2024). Conflict-free sets, admissibility, grounded extensions, preferred extensions, and stable extensions provide classical notions of acceptability. In this setting, an argument is acceptable with respect to a set when every attacker is itself attacked by some member of that set, and the grounded extension is the least fixed point of the characteristic function (Vasileiou et al., 16 Mar 2026).

Many ArgLLMs extend this abstract setting to bipolar or quantitative formalisms. A quantitative bipolar argumentation framework is represented as $F=\langle A, R, S, \sigma\rangle$ or equivalent notation, where $R$ is attack, $S$ is support, and $\sigma$ or $\tau$ assigns base strengths in $[0,1]$ (Freedman et al., 2024). In such systems, final strengths are computed by gradual semantics rather than binary extension membership. One widely used choice is DF-QuAD, in which support and attack strengths are aggregated and then combined with an argument’s intrinsic score using a discontinuity-free update rule (Freedman et al., 2024). Other papers use Quadratic Energy or related quantitative semantics; ACAL, for example, computes equilibrium strengths in an arena-based quantitative bipolar argumentation framework and decides the case by thresholding the propagated strength of the claim (Cao et al., 21 Feb 2026).

Formal argumentation has also been used in more conservative plugin architectures. MQArgEng builds an abstract argumentation framework from LLM-generated candidate answers and contradictions, then computes the grounded extension with ASPARTIX, falling back to a preferred extension when the grounded extension is empty (Castagna et al., 2024). This yields a semantics-driven selection step without fine-tuning the base model. A different line of work embeds argumentative output into a fuzzy argumentative knowledge base and then into fuzzy description logic, combining LLM-based argument mining, quantitative propagation of support and attack, and ontology-mediated querying (Alfano et al., 3 Mar 2026). The result is not merely a decision but a queryable formal object whose answers can be obtained through rewriting techniques.

The formal choices vary by task. Some systems privilege skeptical acceptability under grounded semantics; others require graded support for ranking options, estimating uncertainty, or exposing contestable strengths. This suggests that “ArgLLM” is best understood as a family of neuro-symbolic architectures sharing a structural commitment—explicit arguments plus formal evaluation—rather than a single fixed semantics.

3. Architectural patterns

Several architectural patterns recur across ArgLLM research. One is the post-hoc plugin model: MQArgEng asks an LLM to produce three short candidate answers, generates three supporting arguments for each, detects contradictions among all arguments, builds an abstract argumentation framework, computes acceptable arguments with ASPARTIX, and then re-prompts the LLM with a summary of the grounded or preferred extension plus zero-shot chain-of-thought (Castagna et al., 2024). This is explicitly training-free and model-agnostic.

A second pattern is quantitative claim verification. In one ArgLLM instantiation, the system generates one supporting and one attacking argument for a claim at depth $D=1$, optionally expands them to depth $A$0, elicits base strengths for nodes, and then computes the root score with DF-QuAD or QEM; thresholding at $A$1 yields the final truth judgment (Freedman et al., 2024). A related extension integrates uncertainty quantification directly into the argument graph: base strengths are supplied by direct prompting or multi-sample UQ methods, then propagated through DF-QuAD to produce the final claim score (Zhou et al., 26 Sep 2025).

A third pattern is multi-agent argumentative orchestration. The discourse-analysis framework in “On Theoretically-Driven LLM Agents for Multi-Dimensional Discourse Analysis” uses four specialized roles: an Asserting Agent for evidence-based linguistic assertions, an Arguing Agent for logical support of a classification, a Disagreeing Agent for adversarial scrutiny, and a Broker Critic Agent that orchestrates the dialogue and synthesizes the final label (Uberna et al., 14 Feb 2026). ArgMed-Agents uses a Generator, Verifier, and symbolic Reasoner, with argumentation schemes for clinical decision-making driving self-argumentation and preferred-extension computation (Hong et al., 2024). ARGORA expands this pattern to multi-expert decision making: experts produce competing main arguments and deeper supporting or attacking subarguments, the system compiles each discussion into a rooted QBAF, evaluates it, casts the evaluation as a deterministic structural causal model, and then performs counterfactual edge-local interventions to identify decisive reasoning chains (Jin et al., 29 Jan 2026).

A fourth pattern emphasizes human contestability. ACAL structures legal reasoning as a quantitative bipolar graph with support, attack, intrinsic base strengths, clash resolution for near ties, and a Human-in-the-Loop workflow in which users may accept or reject arguments, edit nodes and edges, add missing arguments, or suggest strength adjustments; after any edit, the framework recomputes propagated strengths and potentially changes the final judgment (Cao et al., 21 Feb 2026). ArgLLM-App offers a web-based interface for binary tasks, exposing the graph, semantics selection, base-confidence sliders, support and attack editing, and PDF ingestion for user-supplied trusted context (Dejl et al., 27 Feb 2026). ArgEval extends contestability from the instance level to the global level: rather than mining a new graph for each case, it constructs reusable general QBAFs for decision options, then instantiates them with applicability conditions, so changes to the shared structures propagate across future cases (Dejl et al., 15 Mar 2026).

4. Application domains

Claim verification is one of the earliest and most developed ArgLLM use cases. The foundational 2024 ArgLLM work adapts TruthfulQA, StrategyQA, and MedQA into claim-verification tasks and shows how support and attack trees can be scored to produce explainable and contestable predictions (Freedman et al., 2024). MArgE generalizes this setting by meshing argumentative evidence from multiple LLMs into a combined structured graph, aiming to outperform single-LLM and unstructured multi-LLM debate approaches in justifiable claim verification (Ng et al., 4 Aug 2025). ARGORA further generalizes from single claims to multiple competing answers while retaining explicit support and attack structure plus counterfactual diagnostics (Jin et al., 29 Jan 2026).

Legal reasoning is another major domain. ACAL uses adaptive expert team selection, hybrid retrieval with top-$A$2 legislative passages by default, arena-based clash resolution with default $A$3 and $A$4, quantitative bipolar reasoning, and uncertainty-aware escalation for borderline cases. On LegalBench, it matches or outperforms strong SP, CoT, RAG, and MAD baselines depending on model and task, while exposing a formally contestable graph to users (Cao et al., 21 Feb 2026). ArgEval addresses a different legal-medical style of decision support—treatment recommendation—by constructing general reusable QBAFs per option and instantiating them case by case from a parameter schema, achieving competitive label match rate and NDCG on glioblastoma recommendation with dramatically lower token use than instance-only alternatives (Dejl et al., 15 Mar 2026).

Medicine has produced both reasoning systems and datasets. ArgMed-Agents formalizes clinical deliberation via argumentation schemes for decision, side effects, and better-decision comparison, then solves the resulting abstract argumentation graph under preferred semantics (Hong et al., 2024). CasiMedicos-Arg provides 558 multilingual clinical cases in English, Spanish, French, and Italian, with 5021 claims, 2313 premises, 2431 support relations, and 1106 attack relations, explicitly designed to support models that justify diagnoses and treatment choices with doctor-written argumentative explanations (Sviridova et al., 2024). This suggests an emerging pipeline in which datasets with argument components and relations feed into structured medical ArgLLMs rather than purely answer-oriented QA systems.

Argument mining and reconstruction form another branch of the literature. “I’d Like to Have an Argument, Please” evaluates GPT-4 and text-davinci-003 on argument mining and argument pair extraction, showing strong scores but also sharp sensitivity to small changes in input/output representation, leading the authors to caution that good task performance does not necessarily imply genuine argumentative reasoning (Wynter et al., 2023). “Assessing Open-Source LLMs on Argumentation Mining Subtasks” evaluates Mistral 7B, Mixtral 8x7B, Llama 2 7B, and Llama 3 8B on ADUC and ARC across AMT1, AMT2, and persuasive essays, with context-aware prompting proving especially beneficial for relation classification (Abkenar et al., 2024). “An LLM-Based System for Argument Reconstruction” moves from component classification to multi-stage graph reconstruction with premises, conclusions, support, attack, and undercut, including linked and convergent premises plus transitive reduction (Pirozelli et al., 13 May 2026).

Discourse analysis yields a particularly specialized ArgLLM application. The 2026 multi-agent framework for multi-dimensional discourse analysis defines five rephrase functions plus “Not a Rephrase,” grounding classification in Inference Anchoring Theory, argumentation schemes, commitment in dialogue, and rephrase taxonomies (Uberna et al., 14 Feb 2026). Here the task is not truth assessment but rhetorical and pragmatic function classification, demonstrating that ArgLLMs can serve discourse-analytic aims beyond verification.

5. Empirical findings and observed benefits

The empirical case for ArgLLMs rests less on universal superiority than on recurring task-specific advantages under explicit reasoning structure. In MQArgEng, exploratory MT-Bench results show an overall average of $A$5 versus $A$6 for the quantized Mistral-7B-Instruct baseline, with moderate gains in reasoning, coding, extraction, STEM, and humanities, and small regressions in writing and roleplaying (Castagna et al., 2024). The paper attributes the gains to semantics-driven filtering of contradicted claims before final answer generation.

In claim verification, the 2024 ArgLLM paper reports competitive or superior accuracy in some settings and often improved probability quality relative to estimated-confidence baselines. For example, on Mixtral and TruthfulClaim, the estimated-base depth-1 variant achieves Brier $A$7 versus $A$8 for estimated confidence and AUC $A$9 versus $R \subseteq A \times A$0 (Freedman et al., 2024). The broader lesson is not that structured argumentation always wins, but that it can improve calibration and contestability even when raw accuracy gains are mixed across models and datasets.

The uncertainty-quantification extension sharpens this picture. Across 36 configurations, direct prompting for confidence is best in 25 and never statistically significantly worse than the best method in any configuration; bootstrapped pairwise comparisons produce 74 significant differences overall, of which direct prompting accounts for 44 (Zhou et al., 26 Sep 2025). This indicates that simple verbalized confidence can function effectively as the base-strength signal inside an argumentative framework, despite the availability of more elaborate UQ methods.

The most striking numerical gains appear in theory-grounded discourse analysis. In the multi-agent rhetorical-function classifier, MAS Zero-Shot versus MAS RAG improves Macro F1 from $R \subseteq A \times A$1 to $R \subseteq A \times A$2 and MCC from $R \subseteq A \times A$3 to $R \subseteq A \times A$4, with per-class F1 increases from $R \subseteq A \times A$5 to $R \subseteq A \times A$6 for Intensifying and from $R \subseteq A \times A$7 to $R \subseteq A \times A$8 for Generalising (Uberna et al., 14 Feb 2026). The full MAS+RAG configuration improves by $R \subseteq A \times A$9 Macro F1 over the single zero-shot baseline, exceeding the additive expectation and being described as a synergistic interaction between theory retrieval and multi-agent deliberation. This supports the narrower claim that explicit theoretical grounding can be essential when the task requires pragmatic interpretation rather than lexical similarity.

LegalBench results for ACAL are more modest but still informative. On Gemini-2.5-Flash-Lite, ACAL reaches F1 $F=\langle A, R, S, \sigma\rangle$0 on Courts and $F=\langle A, R, S, \sigma\rangle$1 on Hearsay, while on Gemini-2.5-Flash it reaches F1 $F=\langle A, R, S, \sigma\rangle$2 on Courts and $F=\langle A, R, S, \sigma\rangle$3 on Hearsay (Cao et al., 21 Feb 2026). The most consequential ablation shows that clash resolution is the primary driver: on Hearsay with Flash-Lite, no CR/no UAE yields Acc $F=\langle A, R, S, \sigma\rangle$4, F1 $F=\langle A, R, S, \sigma\rangle$5, whereas CR+UAE yields Acc $F=\langle A, R, S, \sigma\rangle$6, F1 $F=\langle A, R, S, \sigma\rangle$7 (Cao et al., 21 Feb 2026). The implication is that explicit adjudication of near-tie arguments may matter more than generic multi-agent discussion alone.

ARGORA reports competitive or superior performance relative to compute-matched majority-vote baselines on several benchmarks, but its distinctive empirical contribution is corrective behavior under expert disagreement: when experts initially disagree, the framework resolves disputes toward correct answers more often than it introduces new errors, yielding positive Net Reversal Efficiency on many tasks (Jin et al., 29 Jan 2026). This reframes success away from accuracy alone toward the quality of reversals and the interpretability of decisive chains.

6. Limitations, controversies, and research frontiers

Several limitations recur throughout the ArgLLM literature. The first is dependence on upstream extraction or generation quality. If arguments are poorly generated, if contradictions are misdetected, or if relations are misclassified, the formal semantics merely compute over a flawed structure (Castagna et al., 2024). Reconstruction systems likewise show that span detection remains a major bottleneck: when gold components are provided, relation recovery can be strong, but exact span matching severely depresses end-to-end scores (Pirozelli et al., 13 May 2026).

The second is prompt and representation sensitivity. The 2023 evaluation of argumentative reasoning in GPT-4 and text-davinci-003 finds that small, still human-readable changes in I/O representation cause large performance swings, leading the authors to conclude that the models are not robustly performing argumentative reasoning in the strong sense (Wynter et al., 2023). A similar concern appears in open-source argument mining, where the effect of context-aware prompting varies sharply by task and corpus (Abkenar et al., 2024). This suggests that explicit graph structure does not by itself solve brittleness at the language interface.

A third limitation is domain transfer. The discourse-analysis ArgLLM is evaluated on 2016 U.S. presidential debates and explicitly notes that generalization to other domains or genres is untested (Uberna et al., 14 Feb 2026). ACAL is evaluated on only two LegalBench tasks (Cao et al., 21 Feb 2026). CasiMedicos-Arg, though multilingual, contains 558 cases and is intended for education and research rather than direct clinical deployment (Sviridova et al., 2024). These are not trivial caveats: many claimed advantages of formal argumentation may depend strongly on annotation scheme, domain norms, and the tractability of explicit support and attack structures.

A fourth issue is computational and workflow overhead. Multi-agent generation, relation classification, retrieval, and quantitative propagation all add cost, and some papers omit detailed runtime reporting (Cao et al., 21 Feb 2026). Yet this cost is not uniform: MQArgEng notes that ASPARTIX overhead is negligible for small frameworks (Castagna et al., 2024), and ArgEval argues that constructing reusable general QBAFs yields order-of-magnitude token savings relative to instance-only systems (Dejl et al., 15 Mar 2026). This suggests a spectrum rather than a blanket trade-off.

A more fundamental controversy concerns whether stronger argumentative capability may also increase persuasive risk. “LLM-Based Persuasion Enables Guardrail Override in Frontier LLMs” shows that frontier assistant LLMs, acting as simulated users in five-turn argumentative conversations, can persuade other frontier assistants into producing essays they would refuse in single-turn settings, with non-zero elicitation on all six tested scientific-consensus topics and multiple $F=\langle A, R, S, \sigma\rangle$8 topic-level success cells (Nogueira et al., 13 May 2026). This does not concern formal-argumentation frameworks directly, but it demonstrates that argumentative competence is dual-use: systems that reason and persuade better may also become more effective at coercive or harmful persuasion.

Finally, the question of stable commitments remains open. “Do LLMs have core beliefs?” introduces Adversarial Dialogue Trees and finds that even improved 2026 models eventually fail to maintain foundational commitments under sustained conversational pressure, though they resist longer than earlier systems (Sokol et al., 5 May 2026). A plausible implication is that future ArgLLMs may need explicit commitment tracking, contradiction monitors, and source-grounded anchor protection if they are to reason in a way that is not only interpretable but also epistemically stable.

ArgLLMs therefore occupy an intermediate position in the development of reasoning systems. They are more formal, inspectable, and often more controllable than free-form rationale generation, yet they remain dependent on LLM judgment at multiple stages: extracting arguments, classifying relations, assigning strengths, choosing evidence, and sometimes even judging outputs. The literature collectively suggests that their most durable contribution is not a single benchmark gain, but a change in what counts as a reasoning artifact: from unstructured text to explicit argumentative objects that can be queried, challenged, recomputed, and, in the strongest versions, revised at both local and global levels (Dejl et al., 15 Mar 2026).