Integrated Linguistic-Interpersonal Model
- The Integrated Linguistic-Interpersonal Model is a conceptual framework that unifies linguistic form, discourse performance, strategic adaptation, and social appropriateness into a single competence architecture.
- It provides a multi-level, hierarchical approach to understanding both second-language communication and LLM-user interactions by integrating linguistic, pragmatic, strategic, and social dimensions.
- The model offers practical insights for evaluating communication teaching and human-AI interaction through linking communication form, function, adaptive strategy, and socially recognizable outcomes.
The Integrated Linguistic-Interpersonal Model is a multi-level account of communication competence that treats linguistic form, discourse/pragmatic performance, strategic adaptation, and social/intercultural appropriateness as one connected architecture rather than as separate abilities. In the literature summarized here, the term refers primarily to the "conceptual model of communication competence in foreign language (L2) use" (CMCC-L2), originally developed by Bubaš and Kovačić (2019) and repurposed in 2025 to interpret LLM-user interaction. The model is conceptual, hierarchical, and taxonomic rather than mathematical, and the 2025 study argues that, once the second-language teaching context is excluded, it can be reformulated as a "conceptual model of communication competence of LLMs" (Bubas, 21 Sep 2025).
1. Theoretical foundations
CMCC-L2 is presented as the more theoretically rich of the two focal frameworks in the 2025 pilot study, precisely because it does not reduce communication competence either to narrow linguistic form or to isolated interpersonal skills. Instead, it is characterized as an integration of linguistic elements and interpersonal interaction/skill elements of communication competence. In its original setting, it was designed for second-language use and teaching, but the same structure is explicitly argued to be useful for understanding LLM-user interaction (Bubas, 21 Sep 2025).
Its linguistic lineage is explicitly interdisciplinary. On the linguistic side, the background runs from Chomsky’s distinction between competence and performance to Hymes’s expansion of communicative competence beyond grammar, then to Canale and Swain and Canale on grammatical, sociolinguistic, strategic, and discourse competence. The model is also situated against Bachman and Palmer’s communicative language ability framework, especially the distinction between organizational and pragmatic knowledge and the treatment of strategic competence as metacognitive planning and assessment. Celce-Murcia, Dörnyei, and Thurrell are also identified as influential because their model links sociocultural, linguistic, actional, discourse, and strategic competence (Bubas, 21 Sep 2025).
Its interpersonal side is anchored above all in Spitzberg and Cupach. The 2025 paper emphasizes their tripartite base of skills, knowledge, and motivation, along with criteria for judging competent communication such as effectiveness, efficiency, appropriateness, and satisfaction. It also draws on Spitzberg’s macroscopic, mezzoscopic, and microscopic levels of interaction analysis, on computer-mediated communication competence, on Usó-Juan and Martínez-Flor’s framework integrating the four skills, and on broader contextual approaches including Bronfenbrenner’s social ecological framework, Ting-Toomey’s intercultural framing, and communication accommodation theory as discussed by Zhang. The result is a model whose theoretical ambition is to connect language structure, interactional performance, adaptation, and social consequence within one scheme (Bubas, 21 Sep 2025).
2. Conceptual structure
The formal conceptual structure of CMCC-L2 is given in Table 1 and Figure 1 of the 2025 paper. Table 1 presents a four-level hierarchy, each tied to a competence type.
| Level | Competence type | Description |
|---|---|---|
| Micro | Linguistic competence | “lexicon, phonology, orthography, morphology, syntax, sentence sequencing” |
| Mezzo | Pragmatic/action/discourse competence | “performing and interpreting concise speech acts, monologue, and dialogue according to participant and situational variables” |
| Macro | Strategic/adaptive competence | “knowledge and use of interaction strategies, learning to adapt and advance in competence, utilization of specific skills to enhance ability and overcome barriers” |
| Supra | Social/intercultural competence | “utilization of knowledge, social/cultural cues, and skills to understand the interaction environment and appropriately perform wide-ranging sequences of intentional communication acts” |
Figure 1 extends this hierarchy into a fuller competence architecture by combining potential for competence, enactment of skill(s), and competence confirmation. The potential side derives from Spitzberg’s knowledge-motivation-skills tradition. The listed skills are described as comprehensive and include willingness to communicate, initiation of interaction, listening, self-disclosure, nonverbal sensitivity, self-monitoring, impression management, questioning, empathy, persuasion/assertiveness, conversational skill, expressiveness, social support, and interaction management (Bubas, 21 Sep 2025).
The enactment of skill(s) occurs at the three upper interaction levels: pragmatic/action/discourse, strategic/adaptive, and social/intercultural. Communication competence is then confirmed through “L2 competence confirmation” or social outcomes. The listed outcomes are understanding, appropriateness, influence, coordination, satisfaction, cooperation, efficacy (goal attainment), attractiveness, relationship, inclusion, socialization, learning, competence enhancement, self-development, growth, well-being, and community building. The model is therefore not only a list of abilities; it is a layered pathway from communicative resources and motivations, through enacted communicative performance, to socially recognizable outcomes (Bubas, 21 Sep 2025).
3. Reformulation for LLM-user interaction
The 2025 pilot study operationalized the model by asking three ChatGPT variants and two Gemini variants to read the original 2019 article and reinterpret the framework as if it described the communication competence of LLMs. A key prompt asked whether, with the second-language teaching context excluded, the model could be reformulated as a “conceptual model of communication competence of LLMs.” All tested systems confirmed that they understood this possibility and then produced mappings from the CMCC-L2 structure to LLM-user exchanges (Bubas, 21 Sep 2025).
At the linguistic competence level, the systems treated LLM competence as grammar, lexical choice, syntax, spelling, sentence structure, and semantically coherent output. ChatGPT o4-mini described the analogue as “its mastery of sub-word and word tokens, grammar, spelling, and the statistical patterns of sentence structure in its training data,” and glossed this as “the raw ability to produce well-formed strings of text in the target language.” ChatGPT 4.5 defined this level as the ability to generate “grammatically correct, coherent, contextually appropriate, and semantically meaningful outputs” (Bubas, 21 Sep 2025).
At the discourse/pragmatic/action level, the focus shifted from sentence form to context-sensitive communicative action. The reported interpretations included turn-taking, coherent topic management, speech acts, adaptation to user expectations, and handling ambiguity. Gemini 2.5 Flash summarized this level as the model’s ability to perform and interpret speech acts and sustain coherent monologues and dialogues with attention to user and situational variables. ChatGPT 4.5 emphasized clarifying ambiguous requests, handling misunderstandings, adapting tone and style, and proactively offering useful follow-up suggestions (Bubas, 21 Sep 2025).
At the strategic/adaptive level, the LLMs described a policy-and-planning layer. ChatGPT o3 called it “The policy-and-planning layer: monitors ambiguity, requests clarification, invokes tools, follows safety rules, or changes strategy mid-conversation.” ChatGPT o4-mini likewise emphasized asking clarifying questions, self-correcting after feedback, and using external tools or memory APIs when needed. In this reading, strategic competence becomes the system’s capacity to overcome breakdowns and maintain productive interaction (Bubas, 21 Sep 2025).
At the social/intercultural level, the systems treated competence as alignment with social norms, ethics, cultural expectations, trust, empathy, and inclusion. ChatGPT o3 called it “The alignment layer: modulates tone, mitigates bias, observes cultural taboos, expresses empathy, and sustains trust over extended interactions.” Gemini 2.5 Pro described it as the highest level, where the LLM uses knowledge and skills to understand the broader context of interaction and behave appropriately according to social and ethical norms. The 2025 paper treats this as especially important because it shows that the integrated model can capture not only language production quality but also alignment, appropriateness, and interpersonal calibration (Bubas, 21 Sep 2025).
4. Skill inventory and competence confirmation in LLM interaction
The study also tested how ChatGPT and Gemini interpret the model’s communication-skill inventory, not just its four levels. Table 3 preserves mappings from interpersonal skill labels to LLM-user interaction. Listening was translated into full prompt parsing and attention to all parts of a multi-part user request; ChatGPT o3 called this “textual receptive accuracy.” Self-monitoring was interpreted as internal safety and policy checks, hallucination detection, uncertainty signaling, and adjusting behavior if a response may be inaccurate or biased. Impression management became maintaining a polite, professional, respectful, and authoritative tone. Questioning became clarification-seeking when user goals are ambiguous or underspecified. Empathy was operationalized as supportive, nonjudgmental language and acknowledgment of user feelings. Persuasion/assertiveness appeared as recommending healthier behaviors or gently correcting misconceptions. Conversational skill included turn-taking, discourse markers, topic continuity, and integrating previous conversational details. Expressiveness became varied, engaging style; social support became encouragement, resources, and coping strategies; and interaction management became guiding the user turn by turn through complex tasks (Bubas, 21 Sep 2025).
Two quotations in the paper summarize the interpretive move. ChatGPT o4-mini stated: “Each of these parallels shows that LLMs don’t just generate text—they enact a suite of interpersonal ‘skills’ that mirror those of human communicators, from listening and questioning, to empathy and strategic self-monitoring, thereby building a richly interactive, socially meaningful exchange.” ChatGPT 4.5 similarly stated: “Each communication skill from the original model can thus meaningfully translate into real-world interactions between LLMs and users, offering structured guidance on improving communication effectiveness, enhancing user satisfaction, and facilitating productive outcomes in human-LLM interactions” (Bubas, 21 Sep 2025).
The competence-confirmation side of the model remains central in this reformulation. Because the framework connects potential for competence, enacted competence, and socially recognizable outcomes, it offers a multi-level account in which LLM interaction is not exhausted by textual fluency. The relevant outcomes remain understanding, appropriateness, influence, coordination, satisfaction, cooperation, efficacy, relationship, inclusion, learning, well-being, and community building. This suggests that the model is especially useful when evaluation needs to connect form, function, adaptation, and social consequence rather than treating them as separate benchmarks (Bubas, 21 Sep 2025).
5. Methodological use and empirical findings
The case study around the integrated model was designed to be reproducible in broad outline. Five systems were used in early July 2025: ChatGPT o3, ChatGPT o4-mini, ChatGPT 4.5, Gemini 2.5 Flash, and Gemini 2.5 Pro. The researcher uploaded both the original PDF article by Bubaš and Kovačić (2019) and JPEG images of the relevant figures, because the figures needed to be machine-read and interpreted. Web search, deep research, and temperature manipulation were not used. The procedure unfolded in five stages: first, one figure was uploaded and the model was asked to list the skills in each segment and explain them briefly; second, another figure was used to identify the descriptions and levels of linguistic competence, pragmatic/action/discourse competence, strategic/adaptive competence, and social/intercultural competence; third, the full PDF was uploaded and the model was asked to explain the conceptual model of communication competence in a foreign language, including how it incorporates interpersonal communication skills; fourth, the model was explicitly asked whether the L2 context could be removed and the model reformulated for LLM-user interaction; fifth, after confirmation, the model was asked to connect as many communication skills as possible to LLM-user interaction and provide examples (Bubas, 21 Sep 2025).
The comparison was qualitative and interpretive. The outputs selected for the paper’s tables were chosen on the basis of informational value, showing that all examined LLMs were able to produce valuable output, and avoiding unnecessary repetition across models. The criteria were therefore not numerical benchmarks but interpretive ones: whether the systems could read the figures, explain the model correctly, recognize that it could be adapted to LLM-user interaction, and generate meaningful examples linking model elements to LLM behavior (Bubas, 21 Sep 2025).
The main findings were affirmative on all three research questions. For RQ1, all tested advanced ChatGPT and Gemini models could “read” the figures and “interpret” the PDF content. For RQ2, they could use CMCC-L2 to interpret the way LLMs interact with users. For RQ3, CMCC-L2 was useful for eliciting information from advanced LLMs about their communicative behavior. The paper emphasizes broad convergence between ChatGPT and Gemini: both families recognized the four-level structure, mapped it coherently onto LLM communication, treated linguistic competence as textual correctness and coherence, treated pragmatic/discourse competence as contextual speech-act management, treated strategic competence as adaptation and repair, treated social/intercultural competence as ethics, alignment, context sensitivity, and user-oriented appropriateness, and attached the original interpersonal skill labels to concrete LLM-user interaction scenarios. The differences were mainly stylistic. ChatGPT, especially o3 and 4.5, was described as more architecturally framed and system-like, with labels such as “policy-and-planning layer” and “alignment layer,” while Gemini, especially 2.5 Pro and 2.5 Flash, was described as more pedagogically explanatory and more focused on interactional performance from the user-facing side (Bubas, 21 Sep 2025).
6. Related computational interpretations
Several adjacent studies illuminate components that CMCC-L2 organizes within a single competence architecture. "IMBUE: Improving Interpersonal Effectiveness through Simulation and Just-in-time Feedback with Human-LLM Interaction" operationalizes DEAR MAN as an integrated communication-training system that combines linguistic coaching, interpersonal effectiveness theory, emotional regulation support, interactive simulation, and turn-level feedback; it is explicitly described as the first system to focus on communication skills and emotion management simultaneously, to incorporate expert domain knowledge in feedback, and to be grounded in psychology theory (Lin et al., 2024). This suggests a close parallel to CMCC-L2’s refusal to separate utterance-level skill from broader interpersonal functioning.
Other systems isolate additional parts of the same continuum. "Improving Interpersonal Communication by Simulating Audiences with LLMs" links strategy, candidate utterance generation, and audience-dependent consequences through the Explore-Generate-Simulate framework, thereby treating wording as a vehicle for interpersonal goals and simulated social outcomes (Liu et al., 2023). "An LLM-Guided Tutoring System for Social Skills Training" presents GLOSS, whose narrative graph links what a learner says, how that utterance is interpreted socially, and how the interaction trajectory changes as a result, with immediate feedback and delayed visualization of the learner’s path through the graph (Guevarra et al., 16 Jan 2025). "Modeling Interpersonal Linguistic Coordination in Conversations using Word Mover’s Distance" contributes a unified measure of lexical, semantic, and partly syntactic coordination, and reports correlations with therapist empathy, affective behavior in couples, and relationship improvement over therapy (Nasir et al., 2019). These studies do not present CMCC-L2 itself, but they divide into modules what CMCC-L2 treats as one layered competence path.
A different set of papers extends the same integration into social reasoning, values, and cultural analysis. "Language-Informed Synthesis of Rational Agent Models for Grounded Theory-of-Mind Reasoning On-The-Fly" proposes LIRAS, in which language helps build the agent and environment representations over which Bayesian inverse planning is run, so linguistic input directly affects grounded social inference (Ying et al., 20 Jun 2025). "Value-Based LLM Agent Simulation for Mutual Evaluation of Trust and Interpersonal Closeness" provides evidence for a pathway of internal value configuration, dialogue interaction, and relational judgment, with language acting both as medium and moderator (Sakamoto et al., 16 Jul 2025). "Toward Cultural Interpretability: A Linguistic Anthropological Framework for Describing and Evaluating LLMs" shifts the focus toward communicative competence, cultural interpretability, and the three research axes of relativity, variation, and indexicality, thereby extending the linguistic-interpersonal frame into explicitly cultural evaluation (Jones et al., 2024). Taken together, these works suggest that the integrated linguistic-interpersonal model can function as a higher-level organizing vocabulary for research strands that are often studied separately.
7. Limitations, cautions, and future directions
The 2025 paper is explicit about its limitations. It is a pilot study and qualitative. It analyzes the LLMs’ own “views” of communication competence, not independent behavioral tests by human users or expert raters. The outputs are inherently stochastic, so exact replication is not guaranteed. Some model versions used in the study were no longer available by the time of final submission, although the authors report checking similar procedures with newer systems such as ChatGPT 4o, ChatGPT 5, Copilot with Think Deeper, Claude Sonnet 4, and Claude Opus 4.1. The paper also notes that, because of maximum page length, it could not present a more in-depth analysis of the level of understanding of the CMCC-L2 model or the full extent of its potential for explaining LLM-user interaction. Future directions include applying similar methodology to other communication competence theories, including Spitzberg’s computer-mediated communication competence model and intercultural competence models such as those reviewed by Spitzberg and Changnon, and positioning the work at the overlap of Human-AI Interaction and Explainable AI, especially around social cognition, the uncanny valley in HAI, and more user-centered XAI (Bubas, 21 Sep 2025).
A broader conceptual caution comes from work on reflective discourse and mentalization. "The Linguistic Architecture of Reflective Thought: Evaluation of a LLM as a Tool to Isolate the Formal Structure of Mentalization" reports that generated profiles were coherent and clinically interpretable yet characterized by affective neutrality, with especially clear limitations in integrating internal states and external contexts (Epifani et al., 20 Nov 2025). This suggests that formal linguistic organization, even when it is structurally rich, should not be conflated with lived, affective, embodied, or genuinely intersubjective process.
Practical deployment adds another layer of caution. "From Text to Self: Users’ Perceptions of Potential of AI on Interpersonal Communication and Self" reports that AIMC tools were viewed favorably for increased communication confidence, precise expression, and navigating linguistic and cultural barriers, but also that users identified verbosity, unnatural responses, excessive emotional intensity, inauthenticity, and overreliance as persistent concerns. The same study found the tools more suitable in formal relationships than informal ones and more beneficial in high-stakes than low-stakes communication (Fu et al., 2023). This suggests that any integrated linguistic-interpersonal model intended for deployment must address not only competence representation but also context-sensitive acceptability, authorship, and trust.