Paper deep dive
LLM-enabled Social Agents
Önder Gürcan, Moharram Challenger
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 88%
Last extracted: 7/8/2026, 9:08:14 AM
Summary
This paper proposes a conceptual baseline for LLM-enabled social agents, arguing that sociality must be grounded in persona-based role definitions rather than generic language fluency. It outlines a hybrid architecture where LLMs handle interpretive deliberation, complemented by memory, norms, and control mechanisms. The authors emphasize structured persona representations, hybrid deliberation schemes, and synthetic datasets to evaluate social intelligibility and behavioral coherence in multi-agent systems.
Entities (6)
Relation Signals (5)
LLM-enabled Social Agents → aregroundedin → Persona Descriptions
confidence 93% · This paper develops a conceptual baseline for LLM-enabled social agents by arguing that they should be grounded in role definitions operationalized through persona descriptions.
Large Language Models → enable → Multi-Agent Systems
confidence 90% · LLM-enabled multi-agent systems (MAS) are increasingly embedded in everyday life as they expand their presence across a wide range of domains...
Persona Descriptions → complement → Norms
confidence 88% · To support social deliberation, they should be enriched with complementary social structures, including norms, institutional roles, relationship models, and commitments.
LLM-enabled Social Agents → utilize → Hybrid Deliberation
confidence 87% · this should be understood not as purely LLM-only deliberation, but as LLM-centred hybrid deliberation, in which the LLM remains the main interpretive mechanism while other deliberative and control processes are composed around it.
Persona Descriptions → complement → Memory Systems
confidence 85% · memory systems, relationship models, commitment stores, normative structures, and environmental state can maintain dynamic contextual information outside the LLM.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Large Language Models (LLMs) have transformed agent-agent and human-agent interaction by enabling software, physical, and simulation agents to communicate and deliberate through natural language. Yet fluent language use does not by itself yield socially intelligible behaviour. Most current systems remain weakly grounded in roles, norms, intentions, and contextual constraints, limiting their capacity for meaningful participation in social environments. This paper develops a conceptual baseline for LLM-enabled social agents by arguing that they should be grounded in role definitions operationalized through persona descriptions. On this basis, we outline research directions for representation, hybrid control, and evaluation. The paper concludes that persona-based role definitions are a necessary foundation for turning language competence into social behaviour.
Tags
Links
- Source: https://arxiv.org/abs/2605.02335v1
- Canonical: https://arxiv.org/abs/2605.02335v1
Trouble viewing inline? Open PDF directly →
Full Text
39,784 characters extracted from source content.
Expand or collapse full text
July 20261 LLM-enabled Social Agents ̈ Onder G ̈ URCAN a,1 and Moharram CHALLENGER b a NORCE Research AS, Universitetsveien 19, Kristiansand, Norway. E-mail: ongu@norceresearch.no b University of Antwerp and Flanders Make, Belgium. E-mail: moharram.challenger@uantwerpen.be ORCiD ID: ̈ Onder G ̈ urcan https://orcid.org/0000-0001-6982-5658, Moharram Challenger https://orcid.org/0000-0002-5436-6070 Abstract. Large Language Models (LLMs) have transformed agent–agent and hu- man–agent interaction by enabling software, physical, and simulation agents to communicate and deliberate through natural language. Yet fluent language use does not by itself yield socially intelligible behaviour. Most current systems remain weakly grounded in roles, norms, intentions, and contextual constraints, limiting their capacity for meaningful participation in social environments. This paper de- velops a conceptual baseline for LLM-enabled social agents by arguing that they should be grounded in role definitions operationalized through persona descrip- tions. On this basis, we outline research directions for representation, hybrid con- trol, and evaluation. The paper concludes that persona-based role definitions are a necessary foundation for turning language competence into social behaviour. Keywords. Human-like agents, Persona modeling, Multi-agent systems, Socially- aware behavior 1. Introduction LLM-enabled multi-agent systems (MAS) are increasingly embedded in everyday life as they expand their presence across a wide range of domains including personal as- sistants [1,2,3], autonomous robotics [4,5], autonomous driving, healthcare diagnostics [6,7], supply chains [8], games and social simulations [9,10,11]. Despite this integration, many of these systems are still predominantly designed from a technocentric perspec- tive. Hence, the design and development of these systems often prioritize technical capa- bilities and efficiencies (such as scalability, communication efficiency, and autonomous task decomposition) over social considerations (such as social enactment, fairness, ex- plainability, trust, ethical guidelines and human-centered design), which are critical for real-world deployment and societal acceptance [12,13,10,14]. Recent discourse and em- pirical studies highlight the necessity of balancing technical efficiency with social enact- ment, especially as LLM-MAS are increasingly deployed in domains impacting social welfare, civic engagement, and sensitive decision-making [15,3]. 1 Corresponding Author: ̈ Onder G ̈ urcan, NORCE Research AS, Kristiansand, Norway. Email: ongu@norceresearch.no arXiv:2605.02335v1 [cs.MA] 4 May 2026 2July 2026 Large language models (LLMs) enable natural language understanding [16,17], con- textual interpretation [18,19], flexible response generation [16], and increasingly sophis- ticated forms of reasoning and planning [20,21]. In multi-agent systems, these capa- bilities make it possible to augment more traditional architectures with language-based deliberation, allowing agents to interpret instructions [22,23,24], coordinate with other agents [25,26,27], interact with humans [23,28], and adapt their behaviour to changing contexts [13]. Recent work on LLM-enabled and generative agents further suggests that memory [29], reflection [30,31,32], and persona-related representations [33,34] can sup- port more coherent and context-sensitive behaviour over time. At the same time, existing approaches still often privilege linguistic fluency and task completion [35,36] over per- sistent intention management [35,37], explicit role grounding [38,39], and social action [40,41]. This limitation becomes especially important when considering social agents. A so- cial agent is not merely an entity that communicates in natural language or reacts to ob- servable cues, but one that operates in relation to social roles, norms, expectations, rela- tionships, and situated interaction contexts [42,43,44,45]. In contrast to generalized user modelling or interface-level personalization, social agency requires that social informa- tion be represented within the agent’s internal reasoning process itself, so that perception, interpretation, goal formation, and action selection are shaped by who the agent is, what role it performs, and what constraints and responsibilities follow from that role [46]. From this perspective, social cognition in artificial agents depends not only on interaction capability, but also on structured representations that make behaviour understandable, consistent, and socially meaningful across time and situations [47,48]. The novelty of this paper is conceptual rather than empirical. We argue that persona descriptions should be treated as the representational baseline for LLM-enabled social agents (Section 2), and we use this claim to motivate a research agenda on representation, control, and evaluation (Section 3). 2. The Conceptual Baseline for LLM-enabled Social Agents A conceptual baseline is needed to clarify what should count as an LLM-enabled social agent before discussing architectures, methods, or application domains. Current LLM- enabled agent systems predominantly integrate natural language interaction with task execution, yet they largely treat sociality as an external add-on instead of embedding it as a foundational aspect of the agent’s cognitive model [41,10]. This limits the system’s capacity to interpret intentions, maintain context-sensitive behaviour, and reason about interaction in socially meaningful ways. A stronger baseline is therefore required, one in which sociality is not added after deliberation, but built into the agent definition itself. To justify any baseline for LLM-enabled social agents, it is necessary to consider the criteria such a baseline should satisfy. A useful baseline should be interpretable, so that role-related assumptions remain explicit and inspectable; action-guiding, so that it can meaningfully shape agent behaviour rather than merely label it; compatible with LLM-based deliberation, so that it can be naturally incorporated into language-mediated reasoning; persistent across time, so that the agent can maintain behavioural coherence across interactions; supportive of norm integration, so that social expectations, obliga- tions, and constraints can be attached to the role representation; and sufficiently simple to evaluate, so that alternative implementations can be compared empirically. July 20263 From this perspective, an LLM-enabled social agent should not be understood merely as an agent that can communicate fluently in natural language. Linguistic flu- ency is not sufficient for social meaning. A social agent should be able to act in rela- tion to roles, expectations, constraints, and situated interaction contexts [49]. Its deci- sions should not be generated only from immediate prompts or isolated goals, but from an internally meaningful representation of who the agent is, what role it currently per- forms, what boundaries and responsibilities follow from that role, and how its behaviour should remain interpretable over time. Hence, we define an LLM-enabled social agent as a role-enacting agent whose role is operationalized through a persona description that grounds deliberation in traits, values, goals, constraints, relationships, and commitments, and whose context-sensitive behaviour is generated through LLM-centered deliberation, complemented where needed by additional mechanisms for memory, control, and nor- mative regulation. Figure 1. An LLM-enabled Social Agent’s Internal Architecture (improved from [50]) The key conceptual step is to define LLM-enabled social agents through roles rather than generic capability bundles. In social environments, an agent is a situated participant that occupies one or more roles, and these roles shape what it treats as relevant, what it prioritizes, what actions are legitimate, and how its behaviour is interpreted by oth- ers. However, role labels alone are too weak to support social behaviour. A label such as “caregiver” or “assistant” does not by itself specify how the agent should interpret situations, resolve trade-offs, or sustain behavioural consistency across interactions. For this reason, roles should be operationalized through persona descriptions, which make them computationally meaningful for LLMs by specifying traits, values, goals, needs, skills, constraints, resources, boundaries, and interaction tendencies. Note that, persona descriptions are not understood merely as free-form natural-language prompts. Rather, they should be seen as hybrid symbolic-textual representations: partially readable as natural-language descriptions, but organized into explicit fields such as role, stable traits, values, goals, responsibilities, boundaries, dynamic state, norms, relationships, and com- mitments. In more advanced implementations, such persona representations may also be partially updated or dynamically learned as internal models through interaction, memory, and adaptation over time. As illustrated in Figure 1, an LLM-enabled social agent is grounded in a persona- based role representation, while LLM-based meta deliberation remains central but is 4July 2026 complemented by memory, normative structures, and control mechanisms, , following the broader direction of hybrid role-based architectures for LLM-enhanced multi-agent systems [50]. The persona description can be understood as comprising two complemen- tary parts: a Persona Core and a Persona State. The Persona Core contains the relatively stable elements that preserve the agent’s social identity across situations, such as role grounding, enduring traits, values, boundaries, responsibilities, and longer-lived commit- ments. The Persona State contains the more dynamic elements that change through inter- action and context, such as current interpretations, situational priorities, relationship sta- tus, and memory-informed updates. This distinction helps clarify how persona descrip- tions can remain stable enough to preserve behavioural coherence while also adapting to interaction history, situational change, and environmental feedback. This baseline also implies that persona descriptions should be dynamically assem- bled rather than treated as fixed prompts: stable enough to preserve identity and be- havioural coherence, yet flexible enough to adapt to memory, interaction history, situa- tional change, and environmental feedback. Within this baseline, LLMs provide the main interpretive layer: they map persona-rich role representations onto context-sensitive un- derstanding and action. However, this does not make the architecture LLM-only, nor does it require the LLM to internally remember all evolving social context. Instead, memory systems, relationship models, commitment stores, normative structures, and environmen- tal state can maintain dynamic contextual information outside the LLM. At each delib- eration step, the agent can inject a selected, up-to-date representation of this information into the Persona State. The baseline therefore treats long-term social coherence as an ar- chitectural property of the agent, not as a capability delegated entirely to the LLM. This baseline also clarifies the distinction between generic LLM agents and LLM-enabled social agents. A generic LLM agent may answer, plan, retrieve, or converse effectively without a stable social identity. A social agent, by contrast, must enact a role through a persona description and maintain coherence between persona, context, and behaviour. Note that, we do not claim that persona descriptions are sufficient for social agency. Rather, we claim that they provide an appropriate representational baseline for LLM- enabled social agents, because they make roles behaviorally meaningful in a form that LLMs can operationalize. Norms, obligations, relationships, institutional roles, and memory remain indispensable complementary structures. Persona descriptions should therefore not be understood as isolated character sketches. To support social deliberation, they should be enriched with complementary social structures, including norms, insti- tutional roles, relationship models, and commitments. Norms specify what is permitted, expected, or forbidden; institutional roles specify formal responsibilities and authority relations; relationship models capture interactional asymmetries and mutual expectations between agents; and commitments represent explicitly undertaken obligations that per- sist across interactions. In this way, persona descriptions provide the internal behavioural grounding of the role, while these additional structures provide its normative, relational, and institutional embedding. Several alternative baselines could be used to ground social agency, including role labels, normative rule sets, institutional role models, and relational or commitment-based representations (Table 1). These alternatives are valuable, but they serve different repre- sentational functions. Role labels alone remain too abstract to guide situated behavior, since they specify a category without sufficiently constraining its interpretation in con- text. Normative rule sets capture obligations and prohibitions, but they primarily encode July 20265 Table 1. Comparison of alternative conceptual baselines for LLM-enabled social agents. BaselineStrengthLimitation Role labelSimple, compact, and easy to inter- pret Too abstract to guide situated, socially coherent behaviour Norm rulesMakes obligations, permissions, and prohibitions explicit Weak at capturing internal orientation, style, and open-ended behavioural varia- tion Institutional role modelProvides strong formal structure and clear organizational grounding Less suited to flexible, naturalistic, open- ended interaction Relationship / commit- ment model Captures social dependencies, expec- tations, and inter-agent obligations well Requires broader specification and may be cumbersome as a minimal baseline Persona descriptionBehaviorally interpretable for LLMs and expressive enough to ground role enactment Not sufficient on its own without comple- mentary norms, memory, and relational structure external constraints rather than an internally usable behavioral orientation. Institutional role models formalize positions, permissions, and responsibilities within structured orga- nizations, yet they are less effective as a general baseline for open-ended social interac- tion. Relational and commitment-based models represent social dependencies and expec- tations between agents, but they typically require a richer surrounding social specifica- tion and do not by themselves furnish an immediately actionable deliberative perspective. In comparison, persona descriptions provide a more suitable baseline for LLM-based de- liberation because they render roles behaviorally interpretable in a form that LLMs can directly operationalize. A persona can compactly express motivations, dispositions, com- municative style, contextual expectations, and practical priorities, thereby supplying the LLM with a socially meaningful deliberative frame. For this reason, we argue that per- sonas constitute the most appropriate representational foundation, to be complemented by norms, institutional structure, relationships, commitments, and memory. This reveals what the field is currently missing: a shared baseline that makes so- cial roles computationally actionable for LLM-based agents. On that basis, the next sec- tion proposes a concrete agenda shift toward representation, control, and evaluation of persona-grounded social agents. 3. Research Directions Given this conceptual baseline, three research challenges follow: how persona-grounded social agents should be represented (Section 3.1), how role-consistent behaviour should be preserved through LLM-centred but hybrid deliberation (Section 3.2), and how such agents should be evaluated against weaker alternatives (Section 3.3). 3.1. Representation: From Role Labels to Persona-Grounded Social Models A first research direction concerns how persona-grounded role representations should be specified, organized, and updated. As argued in the previous section, role labels alone are too abstract to guide situated behavior. Labels such as teacher, caregiver, or manager indicate a broad social function, but they do not specify how an agent should interpret 6July 2026 priorities, negotiate expectations, manage tensions, or respond to contextually varied sit- uations. Persona descriptions provide a stronger baseline because they make roles be- haviorally interpretable for LLM-based deliberation by encoding relevant dispositions, tendencies, preferences, commitments, boundaries, and styles of interaction. Future work should therefore investigate how persona descriptions can be designed as structured but flexible representations rather than as informal prompt fragments or static character sketches. In particular, persona-grounded social agents require repre- sentational models that preserve coherence across situations while remaining respon- sive to interaction history, institutional context, and environmental change. This raises questions about which dimensions of persona are most consequential for social behav- ior, how persona elements should be organized, which aspects should remain stable, and which should be dynamically updated over time. A social agent must also be grounded in norms [51,52], commitments [53,54], institutional expectations [55], and relational context [56]. The challenge is therefore not only to describe who the agent is in persona terms, but also to represent what the agent owes to others, what social constraints shape its actions, and how its role is situated within a broader social and organizational envi- ronment. Future research should examine how textual persona descriptions can be com- plemented by more explicit representations of social relationships, normative structures, role obligations, and contextual priorities without losing the expressive advantages of LLM-based interpretation. The field also needs synthetic persona datasets as a shared experimental infras- tructure for LLM-enabled social agents. Machine learning advanced rapidly once com- mon datasets made systematic comparison possible [57]; social-agent research remains largely in a pre-benchmark stage, dominated by bespoke prompts and one-off demon- strations. Without shared persona datasets, it is difficult to determine whether appar- ent social competence reflects genuine role grounding or merely prompt-specific perfor- mance. Carefully designed synthetic persona datasets could encode controlled variation in roles, values, commitments, relationships, cultural settings, institutional constraints, and memory traces, thereby enabling rigorous evaluation of consistency, norm handling, social adaptation, and longitudinal coherence. More importantly, they would allow the field to accumulate knowledge: not just whether a system works in one crafted scenario, but which representational ingredients and architectural choices reliably produce social behavior across settings. Synthetic persona datasets should therefore be treated as a foun- dational research priority rather than a secondary tooling issue. Such datasets would also help scale persona-grounded agent research beyond a handful of manually authored examples. They could enable experiments in which one dimension is varied while others are held constant, making it easier to study the contribu- tion of specific persona elements to behavioral coherence, role fidelity, or social appro- priateness. At the same time, the creation of synthetic persona datasets is itself a research challenge. These datasets may encode cultural narrowness, stereotypical assumptions, or artificial regularities that fail to reflect the complexity of real social life. For that reason, future work should address not only how to use such datasets, but also how to construct them transparently, diversify them responsibly, and document their limitations. 3.2. Control: LLM-Centred but Hybrid Deliberation A second research direction concerns deliberation and execution. Within the proposed baseline, LLMs remain central because they provide the interpretive flexibility needed July 20267 for situated social reasoning, contextual response generation, and adaptive interaction. However, they do not by themselves guarantee persistence of commitments, long-term coherence, reliable planning, or stable enactment of persona-grounded roles. If persona descriptions are to function as more than descriptive overlays, future research must in- vestigate how social agents can preserve persona-grounded behavior throughout deliber- ation, memory use, and action execution. As suggested by Figure 1, this should be un- derstood not as purely LLM-only deliberation, but as LLM-centred hybrid deliberation, in which the LLM remains the main interpretive mechanism while other deliberative and control processes are composed around it. This suggests the importance of hybrid architectures in which LLM-based deliber- ation is complemented by additional mechanisms for memory, planning, constraint han- dling, and action control [50]. The key issue is not whether symbolic or sub-symbolic approaches should dominate, but how complementary mechanisms can support the con- tinued enactment of persona-grounded roles without displacing the LLM’s central role in interpretation and socially situated reasoning. More specifically, hybrid deliberation should be understood as a compositional deliberative scheme: it can incorporate and co- ordinate multiple forms of deliberation, such as reactive deliberation [58], rule-based or normative deliberation [59], planning-based deliberation [60,61], and utility-based de- liberation, argumentation-based deliberation [62,63], or LLM-based deliberation [64]. In this view, hybrid deliberation is not merely the coexistence of modules, but an orchestra- tion layer that selects, combines, constrains, or sequences different deliberative modes depending on the current context, persona state, and role obligations. Such architectures may therefore include explicit memory systems for preserving commitments and rela- tional history, planning modules for maintaining longer-term consistency, reactive mech- anisms for timely situated response, and constraint mechanisms for enforcing normative or institutional boundaries. One particularly important challenge is the coordination of short-term and long-term deliberation. Social agents must often respond to immediate conversational or situational demands while also remaining faithful to enduring commitments [65], role expectations [66,67], and accumulated interaction history. A persona-grounded agent should therefore not only produce locally plausible responses, but also sustain a recognizable pattern of conduct across time. Future work should examine how architectures can maintain conti- nuity between momentary judgments and longer-term persona-grounded role enactment, especially in settings where social expectations evolve or conflict. Memory is central here. Without mechanisms for preserving relevant experiences, commitments, relational trajectories, and prior decisions, persona-grounded behavior risks collapsing into episodic improvisation. Yet memory should not be treated merely as a storage layer. In social agents, memory plays a constitutive role in sustaining who the agent appears to be across repeated interactions [68,69,70]. This raises questions about what should be remembered, how memories should be selected or summarized, how they interact with persona descriptions, and how agents should balance persistent identity with contextual adaptation. 3.3. Evaluation: Measuring Social Intelligibility A third research direction concerns evaluation. If persona-grounded role definitions are to serve as a meaningful conceptual baseline, they must be evaluated not only as an 8July 2026 appealing design intuition but as a stronger foundation than weaker alternatives such as generic LLM agents, role labels alone, or externally imposed rule sets. Future work should therefore develop evaluation frameworks that go beyond task completion and linguistic fluency to assess whether agents behave in social ways. Relevant evaluation criteria include behavioral coherence across interactions, fi- delity to persona and role, persistence of commitments, sensitivity to relational and nor- mative context, and interpretability of behavior to human users or observers. These di- mensions are especially important in domains where agents are expected to participate in socially meaningful interaction (such as social simulations [10]) rather than merely provide information or execute isolated tasks. Evaluation should also examine whether persona-grounded agents remain coherent under contextual variation, conflicting expec- tations, incomplete information, or extended interaction sequences. Shared synthetic persona datasets would be especially valuable in this context be- cause they could support benchmark-style evaluation [71]. By providing standardized but diverse persona configurations, they would enable systematic comparison across agent architectures, prompting strategies, memory mechanisms, and control frameworks. They would also make it possible to conduct ablation-style studies, for example by compar- ing agents grounded in rich persona descriptions with agents specified only through role labels, or by examining how the addition of commitments, norms, or relationship mod- els affects social behavior under controlled conditions. In this way, synthetic persona datasets would support reproducibility, comparability, and cumulative empirical progress in a field that currently relies too heavily on handcrafted examples and anecdotal demon- strations. Evaluation should also include the study of failure modes. Persona-grounded agents may still drift, overfit to superficial cues, become inconsistent over time, or reproduce biased assumptions embedded in their persona specifications or training materials. For this reason, future work should investigate not only success criteria but also mechanisms for diagnosing representational weakness, persona instability, normative breakdown, and socially unintelligible responses. A mature research programme will require methods for identifying where persona-grounded models fail, why they fail, and how such failures differ from those observed in simpler baselines. Taken together, these research directions suggest that persona-grounded social agency should be pursued not merely as a prompting technique or implementation de- tail, but as a broader design and evaluation paradigm. The field now needs richer rep- resentations, hybrid control mechanisms, synthetic persona datasets, and comparative benchmarks that together can support the systematic development of socially coherent LLM-enabled agents. 4. Conclusion We argued that LLM-enabled social agents should be understood as role-playing agents whose roles are operationalized through persona descriptions and enacted through LLM- centered deliberation, complemented by normative, relational, and operational mecha- nisms. This framing provides a clearer conceptual foundation for designing and eval- uating social agent systems. Future work should test this baseline empirically against alternative role-grounding approaches. July 20269 References [1] He G, Demartini G, Gadiraju U. Plan-Then-Execute: An Empirical Study of User Trust and Team Performance When Using LLM Agents As A Daily Assistant; 2025. . [2] Chan S, Li J, Yao B, Mahmood A, Huang CM, Jimison H, et al. “Mango Mango, How to Let The Lettuce Dry Without A Spinner?”: Exploring User Perceptions of Using An LLM-Based Conversational Assis- tant Toward Cooking Partner. Proceedings of the ACM on Human-Computer Interaction. 2025;9(7). [3] Li G, Wu R, Tan H, Chen G. A Plan Reuse Mechanism for LLM-Driven Agent. Jisuanji Yanjiu yu Fazhan/Computer Research and Development. 2024;61(11):2706 – 2720. [4] Rafique Z, Wasim M, Hussain M, Hussain M, Memon MI. From Prompt to Action: A Comprehensive Review of LLM Autonomous Agents; 2025. . [5] Liang G, Tong Q. LLM-Powered AI Agent Systems and Their Applications in Industry; 2025. p. 463 – 471. [6] Xu X, Sankar R. Large Language Model Agents for Biomedicine: A Comprehensive Review of Meth- ods, Evaluations, Challenges, and Future Directions. Information (Switzerland). 2025;16(10). [7] Yang T, Xiao Y, Bao Z, Hao J, Peng J. The rise and potential opportunities of large language model agents in bioinformatics and biomedicine. Briefings in Bioinformatics. 2025;26(6). [8] Zhao X, Kuang W, Kim YW. Leveraging Multi-Agent System Powered by Large Language Model to Improve Transparency and Reliability in Automated Supply Chain Coordination. Annual Conference of the International Group for Lean Construction, IGLC. 2025;33:681 – 692. [9] G ̈ urcan ̈ O, Falck V, Rousseau MG, Lima L. Towards an llm-powered social digital twinning platform. In: International Conference on Practical Applications of Agents and Multi-Agent Systems. Springer; 2025. p. 118-30. [10] G ̈ urcan O. LLM-Augmented Agent-Based Modelling for Social Simulations: Challenges and Opportu- nities. Frontiers in Artificial Intelligence and Applications. 2024;386:134 – 144. [11] Borah A, Houalla M, Mihalcea R. Mind the (Belief) Gap: Group Identity in the World of LLMs; 2025. p. 18441 – 18463. [12] Chowa S, Alvi R, Rahman S, Rahman MA, Raiaan MAK, Islam MR, et al. From language to action: a review of large language models as autonomous agents and tool users. Artificial Intelligence Review. 2026;59(2). [13] Shaikh SH. LLM-Based Multi-agent Systems: Frameworks, Evaluation, Open Challenges, and Research Frontiers. Communications in Computer and Information Science. 2026;2827 CCIS:149 – 170. [14] Sissodia R, Dwivedi V, Verma T. Advancements in Multi-Agent Large Language Model Systems for Next- Generation AI; 2025. [15] Zhang L, Carter RA, Lim SN. Designing Intelligent Agents for Students With Disabilities: Promot- ing Inclusion and Equity Through the Lens of Cultural-Historical Activity Theory. Journal of Special Education Technology. 2025. [16] Shao M, Basit A, Karri R, Shafique M. Survey of Different Large Language Model Architectures: Trends, Benchmarks, and Challenges. IEEE Access. 2024;12:188664 – 188706. [17] Barbon S, Ceravolo P, Groppe S, Jarrar M, Maghool S, S ` edes F, et al. Are Large Language Models the New Interface for Data Pipelines?; 2024. . [18] Mohamed A, Rashid ME, Shaalan K. In-Context Learning in Large Language Models (LLMs): Mech- anisms, Capabilities, and Implications for Advanced Knowledge Representation and Reasoning. IEEE Access. 2025;13:95574 – 95593. [19] Raiaan MAK, Mukta MSH, Fatema K, Fahad NM, Sakib S, Mim MMJ, et al. A Review on Large Language Models: Architectures, Applications, Taxonomies, Open Issues and Challenges. IEEE Access. 2024;12:26839 – 26874. [20] Zhao Z, Wu Z, Meng Q. Large Language Models and Their Applications in the Military Field; 2025. p. 311 – 319. [21] Jiang Y. The Applications of Large Language Models in Emergency Management; 2024. p. 199 – 202. [22] Xiang B. Comprehensive survey on communication structure adaptive control and collaborative opti- mization for multi-agent systems based on deepseek large language model. vol. 13800; 2025. . [23] Lee KY, Tseng PA, Chen WC. Enhancing Telecom Operation Support Systems with Multi-Agent Large Language Models; 2025. p. 309 – 316. [24] Lim J, Vogel-Heuser B, Kovalenko I. Large Language Model-Enabled Multi-Agent Manufacturing Systems; 2024. p. 3940 – 3946. 10July 2026 [25] Yang L, Li S, Deng A. Dynamic Consensus Communication Mechanism for Large Language Model- Based Multi-Agent Systems. Journal of Signal Processing Systems. 2026;98(1). [26] Gogineni V. LLM-Powered Multi-Agent Systems: A Technical Framework for Collaborative Intelli- gence Through Optimized Knowledge Retrieval and Communication; 2025. p. 452 – 456. [27] Chen JCY, Saha S, Bansal M. RECONCILE: Round-Table Conference Improves Reasoning via Con- sensus among Diverse LLMs. vol. 1; 2024. p. 7066 – 7085. [28] Guo X, Huang K, Liu J, Fan W, Velez N, Wu Q, et al. Embodied LLM Agents Learn to Cooperate in Organized Teams. IEEE Transactions on Computational Social Systems. 2026. [29] Jia Z, Li J, Kang Y, Wang Y, Wu T, Wang Q, et al. The AI Hippocampus: How Far are We From Human Memory? Transactions on Machine Learning Research. 2025. Available from: https://openreview. net/forum?id=Sk7pwmLuAY. [30] Tan Z, Yan J, Hsu IH, Han R, Wang Z, Le LT, et al. In Prospect and Retrospect: Reflective Memory Management for Long-term Personalized Dialogue Agents. vol. 1; 2025. p. 8416 – 8439. [31] Liang X, Tao M, Xia Y, Wang J, Li K, Wang Y, et al. SAGE: Self-evolving Agents with Reflective and Memory-augmented Abilities. Neurocomputing. 2025;647. [32] Ma X, Ge X, Xu H, Fu D, Wang Z, Liu Y, et al. FuseMind: Fusing reflection and prediction elevates agent’s reasoning capabilities. Neurocomputing. 2025;658. [33] Tissaoui A. From prompt to persona: a literature review on LLMs as single cognitive agents. Journal of Ambient Intelligence and Humanized Computing. 2026;17(1):205 – 221. [34] Park JS, O’Brien J, Cai CJ, Morris MR, Liang P, Bernstein MS. Generative agents: Interactive simulacra of human behavior. In: Proceedings of the 36th annual acm symposium on user interface software and technology; 2023. p. 1-22. [35] Zhang L, Li Q, Wang Y, Zhao J. Getting LLM to think and act like a human being: logical path reasoning and replanning; 2025. p. 301 – 308. [36] Wu Y, Li D, Chen Y, Jiang R, Zou HP, Huang WC, et al. Multi-Agent Autonomous Driving Systems with Large Language Models: A Survey of Recent Advances, Resources, and Future Directions; 2025. p. 12756 – 12773. [37] Savarimuthu BTR, Ranathunga S, Cranefield S. Harnessing the Power of LLMs for Normative Reason- ing in MASs. Lecture Notes in Computer Science. 2025;15398 LNAI:132 – 145. [38] Takagi H, Moriya S, Sato T, Nagao M, Higuchi K. A Framework for Efficient Development and Debug- ging of Role-Playing Agents with Large Language Models; 2025. p. 70 – 88. [39] Perera R, Basnayake A, Wickramasinghe M. Auto-scaling LLM-based multi-agent systems through dynamic integration of agents. Frontiers in Artificial Intelligence. 2025;8. [40] Munz LPM, Neef C, Preusser I, Richert A. The Double-Edged Sword: Exploring Older Adults’ Inter- action and Imagination with an LLM-Enhanced Health Agent; 2026. p. 95 – 104. [41] Mou X, Liang J, Lin J, Zhang X, Liu X, Yang S, et al. AgentSense: Benchmarking Social Intelligence of Language Agents through Interactive Scenarios. vol. 1; 2025. p. 4975 – 5001. [42] Foo N. Social agents. Lecture Notes in Computer Science (including subseries Lecture Notes in Artifi- cial Intelligence and Lecture Notes in Bioinformatics). 2007;4830 LNAI:14. [43] Jackson RB, Williams T. A Theory of Social Agency for Human-Robot Interaction. Frontiers in Robotics and AI. 2021;8. [44] Governatori G, Rotolo A. BIO logical agents: Norms, beliefs, intentions in defeasible logic. Au- tonomous Agents and Multi-Agent Systems. 2008;17(1):36 – 69. [45] Fasli M. Accounting for social order in multi-agent systems: Preliminary report; 2004. p. 204 – 210. [46] Dignum F, van Linder B. Modelling social agents: Communication as action. Lecture Notes in Com- puter Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioin- formatics). 2015;1193:205 – 218. [47] Hortensius R, Cross ES. From automata to animate beings: The scope and limits of attributing socialness to artificial agents. Annals of the New York Academy of Sciences. 2018;1426(1):93 – 110. [48] Lukasik A, Gut A. From robots to chatbots: unveiling the dynamics of human-AI interaction. Frontiers in Psychology. 2025;16. [49] Sanders MJ. Social role as one explanatory link between individual and organizational levels in meso- ergonomic frameworks. vol. 2014-January; 2014. p. 1400 – 1404. [50] Asici TZ, G ̈ urcan ̈ O, Kardas G. A Hybrid Role-Based Reference Architecture for LLM-Enhanced Multi- Agent Systems. In: Proceedings of the International Workshop on Engineering Multi-Agent Systems (EMAS 2026); 2026. To appear. July 202611 [51] Viana M, Alencar P, Guimar ̃ aes E, Cunha F, Cowan D, Lucena C. JSAN: A framework to implement normative agents. vol. 2015-January; 2015. p. 660 – 665. [52] Dos Santos Neto BF, da Silva VT, De Lucena CJP. An architectural model for autonomous normative agents. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). 2012;7589:152 – 161. [53] Fern ́ andez-Castro V, Pacherie E. Commitments and the sense of joint agency. Mind and Language. 2023;38(3):889 – 906. [54] Carabelea C, Boissier O. Coordinating Agents in Organizations Using Social Commitments. Electronic Notes in Theoretical Computer Science. 2006;150(3 SPEC. ISS.):73 – 91. [55] Fasli M. On the relationship between roles and power: Preliminary report. vol. 1; 2006. p. 313 – 318. [56] Dimas J, Lopes P, Prada R. One for all, all for one: Agents with social identities; 2013. p. 2195 – 2200. [57] Nguyen G, Dlugolinsky S, Bobak M, Tran V, Lopez Garc ́ ıa A, Heredia I, et al. Machine Learning and Deep Learning frameworks and libraries for large-scale data mining: a survey. Artificial Intelligence Review. 2019;52(1):77 – 124. [58] Navarro M, Heras S, Juli ́ an V. Guidelines to apply CBR in real-time multi-agent systems. Journal of Physical Agents. 2009;3(3):39 – 47. [59] Chien SA, Gervasio MT, DeJong GF. On Becoming Decreasingly Reactive: Learning to Deliberate Minimally; 1991. p. 288 – 292. [60] Alford R, Shivashankar V, Roberts M, Frank J, Aha DW. Hierarchical planning: Relating task and goal decomposition with task sharing. vol. 2016-January; 2016. p. 3022 – 3028. [61] Debenham J. Information-based planning and strategies. IFIP International Federation for Information Processing. 2008;276:45 – 54. [62] Kok EM, Meyer JJC, Prakken H, Vreeswijk GAW. A formal argumentation framework for deliber- ation dialogues. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). 2011;6614 LNAI:31 – 48. [63] Tang Y, Parsons S. Argumentation-based dialogues for deliberation; 2005. p. 683 – 690. [64] Yu J, Ding Y, Sato H. DynTaskMAS: A Dynamic Task Graph-driven Framework for Asynchronous and Parallel LLM-based Multi-Agent Systems. vol. 35; 2025. p. 288 – 296. [65] Flores RA, Kremer RC. A pragmatic approach to build conversation protocols using social commit- ments. vol. 3; 2004. p. 1242 – 1243. [66] Galland L, Pelachaud C, Pecune F. Adapting conversational strategies in information-giving human- agent interaction. Frontiers in Artificial Intelligence. 2022;5. [67] Ghosh D, Faik I. Practical empathy: The duality of social and transactional roles of conversational agents in giving health advice; 2021. . [68] Campos J, Kennedy J, Lehman JF. Challenges in exploiting conversational memory in human-agent interaction. vol. 3; 2018. p. 1649 – 1657. [69] Richards D, Bransky K. ForgetMeNot: What and how users expect intelligent virtual agents to re- call and forget personal conversational content. International Journal of Human Computer Studies. 2014;72(5):460 – 476. [70] Aylett R, Kriegel M, Wallace I, Segura E, Mercurio J, Nylander S. Memory and the design of migrating virtual agents. vol. 2; 2013. p. 1311 – 1312. [71] Shaheen H, Marimuthu P, Rashmi D. The Practical Approaches of Datasets in Machine Learning. Lecture Notes in Networks and Systems. 2024;1032 LNNS:221 – 227.