This paper presents our research on Neurosymbolic Argumentative Agents, which bridge a symbolic argumentation framework, grounded in argumentation schemes, with the capabilities of Large Language Models (LLMs). LLMs serve as essential interfaces for understanding and generating natural language argumentation, enabling key tasks such as translating arguments into computational representations, generating natural language arguments, guided argument mining from unstructured text, and, critically, reconstructing enthymemes by inferring missing components using argumentation scheme structures. This neurosymbolic integration leverages the linguistic lfuency of LLMs to enhance the formal rigor of symbolic reasoning, advancing Human-AI Hybrid Intelligence toward richer, argumentation-based interactions.
The field of computational models of arguments has experienced a significant evolution over the past decades. Early research focused primarily on abstract argumentation formalisms, such as that proposed by Dung [e.g., 1], which provided crucial insights into argument acceptability but lacked the internal structure necessary for detailed representation. This foundational work naturally progressed toward structured argumentation models, which explicitly define the internal composition of arguments and the nature of attack relations [e.g., 2, 3, 4, 5].
We argue that the conceptualization of Argumentation Schemes (AS) is vital for modeling human-like
presumptive reasoning within these structured frameworks [
Our research builds directly on this line of development, particularly through the formulation
and implementation of a computational model of argumentation schemes for Multi-Agent Systems
(MAS) [
However, the traditional symbolic paradigm faces considerable challenges when interfacing with human users, who communicate through the fluidity and ambiguity of Natural Language (NL). Human arguments are often enthymemes, arguments with omitted premises or conclusions, requiring sophisticated inferential capabilities from the listener.
In our previous work, we have pursued to bridge the gap between symbolic argumentative agents and
natural language argumentation by incorporating technologies such as chatbot platforms for argument
understanding [
The emergence of LLMs has, however, opened new horizons for Computational Models of Natural Arguments (CMNA). LLMs, built on architectures such as the Transformer and comprising billions of parameters [17, 18], ofer unparalleled capabilities in Natural Language Understanding (NLU) and generation. This positions LLMs as a promising bridge between the formal rigor of symbolic reasoning and the complexity of human communication. The integration of these two paradigms defines Neurosymbolic Argumentative Agents, a key step toward realizing Hybrid Intelligence (HI) [19], grounded in the argumentative capabilities of both AI agents and humans.
This paper presents our progress from symbolic argumentative agents [
Collectively, these contributions outline the approach required for what we envision as the next generation of neurosymbolic argumentative agents, systems capable of fluently processing natural language while preserving the logical rigor and structure necessary for sophisticated autonomous reasoning.
The proposed symbolic argumentative agents approach operates based on a foundational computational
model of AS specifically modeled for MAS [
The framework centers on representing the general structure of AS [
A key contribution of this symbolic model is its mechanism for handling the implicit information inherent to CQs. The approach preserves the essence of Walton’s methodology by keeping the CQs implicit at the implementation level, thus avoiding the necessity of representing them explicitly as additional premises or complex undercutting arguments.
The revelation of relevant implicit information is achieved through the matching between the constructed argument instance and its underlying reasoning pattern []. An argument ⟨, ⟩ is an acceptable instance of its AS to an agent (with knowledge ∆ ) only if: (i) all premises ∈ are supported or inferred from ∆ (the agent ’s knowledge); and (ii) all associated Critical Questions ∈ are positively answered by (∀ ∈ , ∆ |= ). This structure allows CQs to address crucial factors like criticizing premises, pointing out exceptional situations (e.g., source reliability), and representing contextual conditions for the scheme’s use.
This symbolic framework supports agent capabilities in both reasoning and communication: • Argumentation-Based Reasoning: Agents construct and individually evaluate argument acceptability, and subsequently define the collectively acceptable arguments based on argumentation semantics, considering conflicts (attacks) among diferent arguments. The approach allows for AS of diferent levels of specificity, including chained argumentation schemes. • Argumentation-Based Dialogues: The structure facilitates communication by explicitly addressing the problem of scheme awareness. Dialogue protocols define specific performatives, such as question_scheme and inform_scheme, to allow agents to retrieve the specific reasoning pattern used by an opponent when necessary.
Example 1 (Argumentation Scheme Role to Know) The argumentation scheme Role to Know, denoted as role_to_know, is computationally represented as follows: ( role(Agent,Role), role_to_know(Role,Domain), asserts(Agent,Conclusion), about(Conclusion,Domain) ⇒ Conclusion )[as(role_to_know)] with the argumentation scheme name sn = role_to_know, the conclusion C = Conclusion, and premises P = { role(Agent,Role), role_to_know(Role,Domain), asserts(Agent,Conclusion), about(Conclusion,Domain)}.
The associated critical questions CQ are as follows: • role_to_know(Role,Conclusion)[as(role_to_know)]. • reliable(Agent)[as(role_to_know)]. • asserts(Agent,Conclusion)[as(role_to_know)] • role(Agent,Role)[as(role_to_know)].
As an example of instantiating the Role to Know scheme, consider a scenario in which an agent knows that mary plays the role of a doctor – role(mary, doctor). Furthermore, the agent is aware that doctors are knowledgeable about cancer – role_to_know(doctor, cancer).
Suppose mary asserts that “smoking causes cancer” – asserts(mary, causes(smoking, cancer)) – and the agent recognizes that the causes of cancer pertain to the domain of cancer – about(causes(smoking, cancer), cancer). Based on this information, the agent can instantiate the argumentation scheme as follows: ( role(mary,doctor), role_to_know(doctor,cancer), asserts(mary,causes(smoking,cancer)), about(causes(smoking,cancer),cancer) ⇒ causes(smoking,cancer) )[as(role_to_know)]
The tions. as reliable(mary)[as(role_to_know)], agent
can For instance, then it automatically associate the corresponding critical quesmay consider
whether according mary is a reliable
doctor, to the unification function expressed = Role ↦→ doctor, Agent ↦→ mary, Domain ↦→ cancer, Conclusion ↦→ causes(smoking, cancer). 3. Towards a Bidirectional Interface between Natural Language and
The necessity to integrate human dialogue into agent reasoning systems is crucial for the development
of HI [19, 20, 21]. Our initial work and subsequent research focused on creating an efective interface
between natural language used by humans and the computational (symbolic) representation of
arguments employed by MAS [
We developed a method to translate arguments from a computational representation to natural
language using natural language templates associated with AS [15]. The main objective was to enable
agents to explain their reasoning and decision-making to human users by translating computational
arguments into human-readable NL arguments [
Example 3 (Natural Language Template) An example of a natural language template for the argumentation scheme Role to Know is presented below: we should believe that <Conc>."⟩[as(role_to_know)] ⟨“ <Agent> is a <Role>, and <Role>s know about <Domain>. <Agent> asserts that <Conc>; therefore,
Following the Example 2, this template is instantiated as follows: should believe that smoking causes cancer."⟩[as(role_to_know)] ⟨“Mary is a doctor, and doctors know about cancer. Mary asserts that smoking causes cancer; therefore, we
On the input side, we investigated the the use of chatbot technologies (specifically the Rasa framework)
and its NLU module to classify arguments in NL according to the AS used to instantiate them [
Computational Argument
Argumentation Schemes classifier
LLM
Argumentation Schemes
Examples Natural Language Argument
+ Argumentation Scheme
+ Examples necessary for agent reasoning (symbolic) [20, 21]. This capability is crucial for enabling agents to reason over human input.
Later, we proposed the use LLMs to perform this translation from NL arguments into computational, symbolic arguments [20]. The approach is grounded in the use of AS to classify the input argument, providing crucial context and structure to the LLM for the symbolic translation task [20]. By leveraging a Retrieval Augmented Generation (RAG) methodology, the LLM is guided to instantiate variables and predicates accurately into the required computational representation [20]. Figure 1 ilustrates an overview of the proposed approach. The evaluation demonstrated that LLMs eficiently translate simple argument structures and improve performance on complex arguments when provided with increased context (more examples) [20].
This neurosymbolic approach also address the reconstruction of enthymemes, which involves combining the generative capabilities of LLMs with the precise, structured guidance of AS [21]. This is essential because human communication frequently relies on implicit information and enthymemes [21]. In this framework, the LLM is guided by the AS structure to perform inference, reconstructing the missing components (premises or conclusions) and subsequently generating a complete computational representation of the intended argument [21]. This capability, enabled by the LLMs’ semantic understanding and commonsense knowledge, successfully bridges the gap between structured symbolic frameworks and informal natural language expressions, laying the groundwork for sophisticated human-agent interaction within HI systems [21]. The ultimate goal is to enable distributed AI systems to interact meaningfully with human users through shared argumentative structures [21].
Another direction of our research is to leverage the NLU capabilities of LLMs, guided by the structural
patterns of AS, to perform argument mining from unstructured texts (e.g., blog posts or documents) [21].
This approach supports the identification and extraction of argumentative components embedded
within extended, non-argumentative discourse, thereby integrating sub-symbolic processes (extraction
and translation via LLMs) with the agent’s symbolic reasoning engine [
The approach provides two main novel contributions: (i) demonstrating how argument mining is performed using AS structures to guide LLMs in argument extraction, and (ii) showing how the extracted arguments are translated from natural language into formal symbolic representations. The method operates as an LLM-based argument extraction pipeline that processes input text, searches for argumentative sentences, performs AS classification, and translates the findings into a computational form.
1. Text Pre-processing: The input document (e.g., everyday discourse or formal documents) is pre-processed by an initial LLM instance.
Document 1 Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
Document 2 Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
aliqua.
Lorem ipsum doLloresmit iapmsuemt, cdonlosresciteatumr eatd,icpoisnicsiencgt etluitr, asdeidpidsoicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
e usmo tempor incididunt ut labore et dolore magna Lorem ipsum doaloiqrusait.Laomr emt, ciposnusmecdteotluor asditipaimsiecitn,gcoenlist,escetdtudroadipisicing elit, eiusmod tempor incididunt ut labore et dolore magna aliqua.
sed o eiusmod tempor incididunt ut l bore et dolore magna Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
Lorem psum dol sit amet, consectetur adipisicing elit, sed do Lorem ipsum doelioursmsitoadmt emt,pcornisne cidt eidturntaudtiplaisbicoirneg etlidt,osloerdedmoagna aliqua. eiusmod tempor incididunt ut labore et dolore magna aliqua.
Lorem psum dol sit amet, consectetur adipisicing elit, sed do
Document n Lorem ipsum doelioursmsitoadmt emt,pcornisne cidt eidturnatduitpliasbicoinreg etlidt,osloerdedmoagna aliqua. eiusmod tempor incididunt ut labore et dolore magna aliqua.
Lorem ipsum doLloresmit iapmsuemt, cdonlosresciteatumr eatd,icpoisnicsiencgt etluitr, asdeidpidsoicing elit, sed do
aliqua. eiusmod tempor incididunt ut labore et dolore magna aliqua.
e usmo tempor incididunt ut labore et dolore magna Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do
L em ipsum daloiqrusai.Laomemt, conusmecdtotuor aditipaimsiecitn,gcoenlist,escetdtudroadipisicing elit, Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do eiusmod tempoer iunscmidoidutnetmupt olarbinocreideidtudnotlourtelamboargeneatadlioqluoare. magna aliqua.
sed o eiusmod tempor ncididunt ut l bore et dolore magna eiusmod tempor incididunt ut labore et dolore magna aliqua.
Lorem psum dol sit amet, consectetur adipisicing elit, sed do Lorem ipsum doelioursmsitoadmt emt,pcornisne cidt eidturnatduitpliasbicoinreg etlidt,osloerdedmoagna aliqua.
Lor m psum dol sit amet, consectetur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
Lorem psum doeliusmsitoadmt emt,pcornisne cidt eidturnatditpliasbicoinreg etlidt,osloerdedmoagna aliqua. eiusmod tempor incididunt ut labore et dolore magna aliqua.
Lorem psum dol sit amet, consectetur adipisicing elit, sed do Lorem ipsum doelioursmsitoadmt emt,pcornisne cidt eidturnatduitpliasbicoinreg etlidt,osloerdedmoagna aliqua. eiusmod tempor incididunt ut labore et dolore magna aliqua.
Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do
L em ipsum d lor si amet, consect tur adipisicing elit, sed do eiusmod tempoer iunscmidoidutnetmupt olarbinocreideidtudnotlourtelamboargeneatadlioqluoare. magna aliqua.
Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
Lorem psum dol sit amet, consectetur adipisicing elit, sed do Lorem ipsum doelioursmsitoadmt emt,pcornisne cidt eidturnatduitpliasbicoinreg etlidt,osloerdedmoagna aliqua. eiusmod tempor incididunt ut labore et dolore magna aliqua.
Lorem psum dol sit amet, consectetur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
Argumentation
Schemes <{predicate1(A),predicate2(A,B),predicate3(C)},predicate(B)> <{predicate8(X),predicate9(Y,X)},predicate(Y)>
...
<{predicate4(D,E),predicate5(E,F)},predicate(F)> LLMs
LLM-Based Argument Mining
Component
Document 1 Lorem ipsum dDolorosictaumemt,coensnecttet2ur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
L em ps m d l si amet, con ect tur adipisicing elit, sed do Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do
eiusmod t mp Dri oidcidunt mut laebonret tndolore magna eiusmod tempor incididunt ut labore et dolore magna aliqua.
al qua.Lorem ipsum dolor sit a et, consectetur adipisicing elit, sed do eiusmod t mpor inci dunt ut labor t lore magna Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do
al qua. eiusmod tempor incididunt ut labore et dolore magna aliqua.
Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do
L em ps m d l si amet, con ect tur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
qua. eiusmod tempor incididunt ut labore et dolore magna aliqua.
L em ps m d l si amet, con ect tur adipisicing elit, sed do
L em psum d lor si amet, con ect tur adipisicing elit, sed do Lorem ipsum dolor sit amet, ceoinusemcotedturmadpiprisicindgidelint, suet dladbore t dolore magna
e usmo tempor incididunt ut labore et dolore magna aliqua. eiusmod tempor incididunt ut labqourae.Leotrdeomloirpesmumagdnoaloarlisqiuta. et, consectetur adipisicing elit, Lorem ipsum dolor sit amet, csoendsedcotetiursamdoipdisticminpgoerlint,csiedudnot ut labor t lore magna
L em ps m d l si amet, con ect tur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
eiusmod t mp r i idid nt ut labore t dolore magna aliqua.
Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do
L em ps m d l si amet, con ect tur adipisicing elit, sed do eiusmod tempor incididunt ut Llaobroerme eptsdomlodre lmasginamaleiqt,ucao.n ect tur adipisicing elit, sed do e usmo tempor incididunt ut labore et dolore magna aliqua.
eiusmod tempor incididunt ut labore et dolore magna aliqua.
L em psum d lor si amet, con ect tur adipisicing elit, sed do
L em psum d lor si amet, con ect tur adipisicing elit, sed do Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do
e u mo empor n didun ut lab re et dolore magna aliqua.
e usmo tempor incididunt ut labore et dolore magna aliqua. eiusmod tempor incididunt ut labore et dolore magna aliqua.
L em ps m d l si amet, con ect tur adipisicing elit, sed do Lorem ipsum dolor sit amet, cLonrsemcteptsur madipislicisnigaemlite,ts,ecdondoect tur adipisicing elit, sed do
eiusmod t mp r i idid nt ut labore t dolore magna aliqua. eiusmod tempor incididunt ut elaiubsomreoedt domloprer miagidnida anlitqutal. bore t dolore magna aliqua.
L em ps m d l si amet, con ect tur adipisicing elit, sed do Lorem ipsum dolor sit amet, cLonrsemcteptsur madipislicisnigaemlite,ts,ecdondoect tur adipisicing elit, sed do
e usmo t mp r i idid nt ut labore t dolore magna aliqua. eiusmod tempor incididunt ut elaubsomreoet deomloproer minacgidnidauanltiqutal. bore et dolore magna aliqua.
L em psum d lor si amet, con ect tur adipisicing elit, sed do Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do
e usmo tempor incididunt ut labore et dolore magna aliqua. eiusmod tempor incididunt ut labore et dolore magna aliqua.
L em ps m d l si amet, con ect tur adipisicing elit, sed do Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do
eiusmod t mp r i idid nt ut labore t dolore magna aliqua. eiusmod tempor incididunt ut labore et dolore magna aliqua.
L em ps m d l si amet, con ect tur adipisicing elit, sed do Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do
e usmo t mp r i idid nt ut labore t dolore magna aliqua. eiusmod tempor incididunt ut labore et dolore magna aliqua.
Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
Lorem ipsum dolor sit amet, consectetur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Figure 2: Overview of the Argument Mining Approach
[21].
The pre-processed text, along with a predefined set of AS and a prompt, is provided to a second LLM instance.
The prompt, defining the task, provides AS examples (including natural language description and computational representation), and instructs the model to match text excerpts
with the appropriate scheme.
The LLM is specifically tasked with ifnding (or inferring) missing premises or conclusions, i.e., enthymemes, based on the contextual information and the structure defined by the AS. This capability allows the construction of complete argument structures from naturally occurring, often incomplete, discourse.
The
LLM performs instantiation by selecting the appropriate scheme (e.g.,
and mapping variables (e.g., Agent, Role, Domain, Conclusion ) to specific terms and predicates found or inferred from the text.
into a formal computational representation compatible
with existing computational argumentation frameworks used by agents, such as our approach [
2.
This symbolic output is then processed by neurosymbolic argumentative agents to perform tasks such as evaluating argument acceptability by checking associated CQs or determining relationships (support, attack) between arguments in a multi-agent system.
This capability enables AI agents to directly interpret, reason about, and communicate arguments extracted from diverse sources, leveraging them for advanced reasoning and interaction. This strategy lays the foundation for distributed AI systems capable of engaging meaningfully with human users, both synchronously and asynchronously, within
Firstly, we have proposed approaches to encoding and decoding enthymemes within a symbolic representation [22, 23, 24, 25], enabling more eficient argumentation-based dialogue among computational agents.
We then moved to the advanced functionality in which LLMs, guided by AS structures, infer and reconstruct missing premises or conclusions in enthymemes expressed in natural language [26].
represented arguments (e.g., inferring missing contextual premises such as role_to_know(doctor, cancer)), thereby making implicit human reasoning explicit for the AI agent’s reasoning mechanisms.
Example 4 (Natural Language Enthymeme) An example of natural language enthymeme, instantiated of the argumentation scheme Role to Know, is as follows: “Mary asserts that smoking causes cancer; therefore, we should believe that smoking causes cancer” In contrast to the natural language argument generated in Example 3, the following argument omits the premises “Mary is a doctor” and “doctors know about cancer.” However, the proposed approach is capable of eficiently reconstructing the complete argument by using the AS as a guide. It does so by aligning variables across diferent predicates and leveraging the general contextual knowledge encoded within the LLM. For this particular example, the system generates an output equivalent to the computational argument presented at the end of Example 1.
Enthymemes, defined formally as an incomplete argument ⟨′, ′⟩ where components (∆ ) are omitted from the full argument ⟨, ⟩ such that (′ ∪ ′) = (( ∪ ) ∖ ∆) , represent a complex yet natural form of human communication. The challenge for intelligent agents in hybrid systems is the decoding process, reconstructing the missing content based on implicit shared knowledge or context.
Our methodology proposes a neurosymbolic interface that leverages the LLMs’ advanced capabilities in NLU to handle ambiguity and context-dependency inherent in enthymematic discourse, combining them with the structured reasoning provided by AS. AS serve as abstract, structuring rules, acting as templates to guide the reconstruction process and enable agents to access the implicit knowledge embedded within reasoning patterns.
The core of the approach is an LLM-based reconstruction component that processes a natural language enthymeme ⟨′, ′⟩ and outputs a complete computational argument ⟨, ⟩ .
1. AS Classification: The LLM first classifies the input enthymeme based on a predefined set of AS
(∆ ) to identify the underlying reasoning pattern (e.g., Argument from Role to Know). This is
guided by prompt engineering that includes constructive examples of AS in both natural language
and computational form.
2. Inference and Reconstruction: Guided by the classified AS structure, the LLM infers the
missing components (∆ ), premises or conclusions, that were omitted in the original enthymeme,
making the implicit human reasoning explicit. The LLM utilizes its extensive linguistic and world
knowledge to hypothesize these missing components, efectively modeling what is "taken for
granted" in the communication context.
3. Symbolic Instantiation: The LLM performs instantiation by mapping the variables in the
selected scheme to specific terms and predicates found in, or inferred from, the text. For
instance, in Example 4, the LLM infers the missing premises role_to_know(doctor, cancer)
and about(causes(smoking, cancer), cancer) to complete the structure of the Role to Know
scheme.
4. Computational Output Generation: The final output is the complete argument represented in
a computational form, ensuring compatibility with computational argumentation frameworks
used, for example [
This reconstruction process enhances agent capabilities, allowing them to comprehend human communication more fully and make informed decisions based on the enriched, structured input. Our evaluation, in [26], demonstrates that LLMs achieve good accuracy in classifying enthymemes and efectively reconstruct arguments, even with significant implicitness (e.g., 50% of missing premises/conclusion), provided they are anchored to these structured AS guides.
The future of argumentative agents remains uncertain. While LLMs evolve and expand their applicability across diverse domains, they continue to transform our understanding of what machines can achieve in communication, reasoning, and collaboration. LLMs have already demonstrated extraordinary capabilities in comprehending and generating natural language, supporting an ever-growing range of cognitive and decision-making tasks that were traditionally handled by symbolic AI [28]. Their lfexibility, contextual awareness, and apparent ability to reason across vast semantic spaces position them as powerful tools for implementing the next generation of argumentation-based systems.
However, while these sub-symbolic systems ofer unprecedented performance in language processing and knowledge synthesis, their inherent opacity and lack of formal guarantees challenge their suitability for domains where reasoning must be transparent, accountable, and logically sound [29]. Argumentation, as a field rooted in logical structure and epistemic justification, demands such rigor. This tension reveals a key insight: the future of argumentative agents might not be defined by the dominance of either symbolic or sub-symbolic paradigms alone, but by their synthesis.
Hybrid, neurosymbolic approaches emerge as a compelling middle ground, combining the interpretability, verifiability, and formal soundness of symbolic reasoning with the expressiveness, adaptability, and contextual fluency of neural models. In this hybrid vision, LLMs act not as autonomous reasoners but as perceptual and linguistic interfaces that ground and enrich the agent’s symbolic reasoning processes [29]. AS provide the structural basis through which LLMs can anchor their generative potential to logical form, ensuring that reconstructed, translated, or mined arguments remain consistent with the agent’s epistemic framework.
Beyond this technical synthesis, hybrid architectures open new horizons for integrating humaninspired reasoning mechanisms, such as Theory of Mind (ToM), the capacity to model and anticipate the beliefs, intentions, and emotions of others. Embedding ToM-like capabilities within argumentative agents allows them to interpret communicative intent, adapt argumentative strategies, and engage in more context-sensitive, cooperative, and persuasive dialogues [25, 22]. By reasoning about the mental states of their interlocutors, these agents can better align their arguments with human expectations, thus advancing both communicative efectiveness and epistemic alignment.
The development of such architectures points toward a new class of Neurosymbolic Argumentative Agents, capable of combining diverse forms of reasoning, deductive, abductive, and presumptive, with mechanisms for perspective-taking and communicative adaptation. These agents could operate in dynamic, open environments, where they interpret, evaluate, and generate arguments in ways that align both with human communication norms and with formal reasoning requirements.
However, achieving this vision will require overcoming significant challenges. Future research must address the alignment of LLM outputs with formal semantics, the incorporation of grounding and truth-maintenance mechanisms, and the establishment of robust evaluation methodologies that assess both argumentative quality and epistemic reliability. Moreover, as neurosymbolic systems become more deeply embedded in human decision-making processes, ethical and epistemological questions regarding responsibility, bias, and interpretability will become central to their design.
In this context, we envision a trajectory where neurosymbolic argumentative agents become foundational components of HI systems, ones that do not merely imitate human reasoning but extend it through structured, transparent collaboration between human and artificial agents. It seems to be a promising direction for developing argumentative agents that are not only powerful but also trustworthy, explainable, and capable of reflective, socially aware reasoning.
In this paper, we provided an overview of our research on argumentative agents, reflecting on the
conceptual and technical evolution of our work from symbolic to neurosymbolic paradigms. Our reflection
began with the foundational computational model of AS for MASs [
Building upon this foundation, we extended our research toward new horizons that explore the intersection between natural language and symbolic reasoning. We incorporated Natural Language Processing (NLP) methods and, more recently, LLMs, to bridge the gap between formal argument representation and human communication. These eforts culminated in neurosymbolic approaches that employ LLMs for argument translation [20], argument mining [21], and enthymeme reconstruction [26], tasks that combine the contextual and linguistic capabilities of neural models with the epistemic precision of symbolic reasoning.
Through these developments, we outlined a trajectory toward Neurosymbolic Argumentative Agents, which we see as a promising horizon for the next generation of research on computational models of natural argument. In such systems, LLMs provide adaptive understanding and generative capacity, while symbolic structures guarantee consistency, interpretability, and accountability. Together, they embody a step toward HI, where human and artificial reasoning processes are integrated through shared argumentative structures.
Our reflection over the past years and our horizon for the future converge on a central insight: progress in computational argumentation depends not only on increasing linguistic or computational capacity but also on preserving the principles of transparency, justification, and reasoning integrity that define argumentation itself. As we look toward the coming decades, we envision hybrid, neurosymbolic argumentative agents that advance these principles, agents capable of reasoning, explaining, and engaging in dialogue with humans in ways that are both intellectually rigorous and deeply aligned with the goals of collaborative, trustworthy AI.
During the preparation of this work, the author(s) used Generative AI in order to grammar, formal tone, and spelling check. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content. [12] M. d. S. Custódio, et al., Rasa4jaca: Uma interface entre sistemas multiagentes e tecnologias chatbots open source (2022). [13] D. Engelmann, J. Damasio, T. Krausburg, O. Borges, M. Colissi, A. R. Panisson, R. H. Bordini, Dial4jaca–a communication interface between multi-agent systems and chatbots, in: International conference on practical applications of agents and multi-agent systems, Springer, 2021, pp. 77–88. [14] A. R. Panisson, D. C. Engelmann, R. H. Bordini, Engineering explainable agents: an argumentationbased approach, in: International Workshop on Engineering Multi-Agent Systems, Springer, 2021, pp. 273–291. [15] C. E. A. Ferreira, D. C. Engelmann, R. H. Bordini, J. L. Carbonera, A. R. Panisson, A knowledge base of argumentation schemes for multi-agent systems, in: International Conference on Enterprise Information Systems (ICEIS), 2024, pp. 1–15. [16] C. E. A. Ferreira, A. R. Panisson, D. C. Engelmann, R. Vieira, V. Mascardi, R. H. Bordini, Explaining semantic reasoning using argumentation, in: International Conference on Practical Applications of Agents and Multi-Agent Systems, Springer, 2022, pp. 153–165. [17] W. X. Zhao, K. Zhou, J. Li, T. Tang, X. Wang, Y. Hou, Y. Min, B. Zhang, J. Zhang, Z. e. a. Dong, A survey of large language models, arXiv preprint arXiv:230318223 (2023). [18] J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, et al., Gpt-4 technical report, arXiv preprint arXiv:230308774 (2023). [19] Z. Akata, D. Balliet, M. de Rijke, F. Dignum, V. Dignum, G. Eiben, A. Fokkens, D. Grossi, K. Hindriks, H. e. a. Hoos, A research agenda for hybrid intelligence: Augmenting human intellect with collaborative, adaptive, responsible, and explainable artificial intelligence, Computer 53 (2020) 18–28. [20] G. Trajano, D. C. Engelmann, R. H. Bordini, S. Sarkadi, J. Mumford, A. R. Panisson, Translating natural language arguments to computational arguments using llms, in: Computational Models of Argument (COMMA), IOS Press, 2024, pp. 289–300. [21] G. Trajano, S. Sarkadi, A. L. Gonçalves, A. R. Panisson, Towards argument mining agents: A neurosymbolic approach based on argumentation schemes and llms, in: IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technology (WI-IAT), 2025. [22] A. R. Panisson, P. McBurney, R. H. Bordini, Towards an enthymeme-based communication framework in multi-agent systems, in: Proceedings of the 19th International Conference on Principles of Knowledge Representation and Reasoning, KR 2022, Haifa, Israel. July 31-August 5, 2022, 2022, Israel., 2022. [23] A. R. Panisson, P. McBurney, R. H. Bordini, Towards an enthymeme-based communication framework, in: 21st International Conference on Autonomous Agents and Multiagent Systems, AAMAS 2022, Auckland, New Zealand, May 9-13, 2022, 2022, Nova Zelândia., 2022. [24] A. R. Panisson, Argumentation schemes and enthymemes in multi-agent systems, in: Proceedings of the 16th Conference on Autonomous Agents and MultiAgent Systems, 2017, pp. 1849–1850. [25] A. R. Panisson, R. H. Bordini, Uttering only what is needed: Enthymemes in multi-agent systems, in: Proceedings of the 16th International Conference on Autonomous Agents & Multiagent Systems (AAMAS-2017), 2017, Brasil., 2017. [26] G. Trajano, S. Hofmann, S. Sarkadi, A. R. Panisson, Toward enthymeme reconstruction: A neurosymbolic interface, in: IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technology (WI-IAT), 2025. [27] A. R. Panisson, M-arguments, in: 2020 IEEE/WIC/ACM International Joint Conference on Web
Intelligence and Intelligent Agent Technology (WI-IAT), IEEE, 2020, pp. 161–168. [28] S. J. Russell, P. Norvig, Artificial intelligence: A modern approach, global edition 4e (2021). [29] U. Nawaz, M. Anees-ur Rahaman, Z. Saeed, A review of neuro-symbolic ai integrating reasoning and learning for advanced cognitive systems, Intelligent Systems with Applications (2025) 200541.