CLT, originating in the late 1970s, counters traditional grammar-translation methods by emphasizing communicative competence, integrating grammatical, sociolinguistic, discourse, and strategic elements (Richards and Rodgers, 2014; Larsen-Freeman and Anderson, 2011). Recent developments underscore hybrid models blending explicit grammar with interactive tasks like role plays to address EFL challenges (Sreehari, 2012; Toro et al., 2019; Dos Santos, 2020). Empirical studies in ESL settings affirm CLT’s efficacy in enhancing motivation and skills (Thamarana, 2015; Alsyed, 2018), while in China, research reveals tensions between grammar dominance and communicative goals (Chen et al., 2024). These findings highlight a need for quantitative examinations of pre-service cognitions in exam-oriented contexts like China to bridge theoretical endorsements with practical adaptations, which this study addresses through SEM modeling of perceptions, attitudes, and beliefs.

Despite endorsement, CLT’s integration in EFL contexts faces systemic barriers, including exam-driven curricula and hierarchical teacher roles (Li, 1998; Gorsuch, 2000; Hu, 2005). In East Asia, educators struggle to reconcile CLT’s learner-centered approach with assessment traditions (Chung and Huang, 2009; Tootkaboni, 2019; Jin and Yoo, 2019). In China, high-stakes exams perpetuate grammar focus, leading to deviations from CLT principles due to proficiency gaps and class sizes (Qi, 2007; Yan, 2015; Chen et al., 2024; Chen et al., 2016). Global studies indicate hybrid implementations where theoretical support coexists with limited uptake (Karimi and Biria, 2017; Alharbi, 2022; Dang and Seals, 2018). Transitioning to teacher cognition, these challenges underscore the cognition-practice gap; this study fills empirical voids by quantifying how pre-service perceptions in high-stakes Chinese EFL settings influence attitudes and beliefs, informing localized reforms.

Teacher cognition encompasses beliefs, perceptions, and attitudes that shape instructional decisions, often rooted in early experiences and resistant to change (Borg, 2003, 2011). In CLT contexts, a cognition-practice gap emerges: educators affirm communicative principles but revert to grammar methods under institutional pressures (Savignon and Wang, 2003; Zhu and Shu, 2017). This gap is contextually mediated, with exam systems prioritizing accuracy over interaction (Qi, 2005; Chen et al., 2016). Recent quantitative studies on attitude-belief discrepancies in Chinese EFL teachers reveal implicit-explicit tensions (Sun et al., 2022; Gong et al., 2022). Attitudes, as evaluative stances, mediate this gap by filtering perceptions into beliefs, per Borg’s (2011) dynamic framework; yet, mediation models remain scarce in pre-service research. This study justifies and tests attitudes’ mediating role via SEM, addressing quantitative gaps in Chinese pre-service EFL cognition dynamics.

Participants were selected through a hybrid approach: convenience sampling from two accessible normal universities in China (key for training secondary English teachers), supplemented by snowball sampling (see section 3.3), with purposive screening applied post-collection to ensure alignment with the target population. These represent a subset of the broader population of pre-service English teachers (English-major undergraduates) across China. Participants comprised 865 English-major undergraduates at two normal universities in China—key institutions training future secondary school English teachers. The sample included 143 males (16.5%) and 722 females (83.5%), reflecting the gender imbalance typical in Chinese English education programs (Shan and Li, 2020). Regarding English learning onset, 67.4% began in primary school, 16.4% pre-primary, 15.4% in middle school, and 0.8% in high school. This distribution enabled subgroup analyses by gender and starting age. A priori power analysis using G*Power (Faul et al., 2009) confirmed the sample size provided > 0.95 power for detecting medium effects in SEM (Cohen’s f² = 0.15, α = 0.05). The sample’s diversity in learning onset allowed exploration of how early exposure influences cognitions, a factor highlighted in recent studies on pre-service teacher development (Yang, 2019).

The questionnaire, adapted from Savignon and Wang (2003) and contextualized for Chinese EFL, included 51 items across four sections: (1) demographics (e.g., gender, English starting age); (2) perceptions of secondary classroom practices (10 items: 5 grammar-focused, 5 communication-focused); (3) attitudes toward practices (10 items: parallel structure); and (4) beliefs about English learning (18 items: 10 grammar-focused, 8 communication-focused). Items used a 5-point Likert scale (1 = strongly disagree to 5 = strongly agree). To enhance validity and balance viewpoints, items were constructed with a mix of positive (e.g., endorsing communicative practices) and negative (e.g., favoring grammar dominance or critiquing interaction) statements across subscales, reducing acquiescence bias. Reverse-scored items were recorded during analysis (see section 3.4) to ensure consistent interpretation. Pilot testing was conducted with 50 similar but independent participants (English-major undergraduates from comparable institutions, not part of the main sample of 865) to ensure the instrument’s clarity and cultural appropriateness, yielding minor revisions. Bilingual back-translation ensured cultural equivalence (Brislin, 1970). Initial reliability checks showed acceptable Cronbach’s α (>0.70 for most subscales), with the grammar-perception subscale at 0.599. Despite this lower value, it was retained due to strong theoretical relevance to CLT contrasts in exam-oriented contexts, as item-total correlations (>0.30) indicated conceptual cohesion without redundant items. Sensitivity tests, excluding borderline items, confirmed consistent SEM results, supporting its inclusion without compromising overall validity. The instrument’s adaptation included items on digital tools to reflect contemporary EFL trends, enhancing relevance to current educational contexts (Guan et al., 2024).

Data were gathered online via Wen Juan Xing (WJX), a secure platform compliant with data protection regulations. The questionnaire link was disseminated through university emails and social media, supplemented by snowball sampling for broader reach. Snowball sampling was used to extend the initial convenience sample, leveraging participants’ networks to increase diversity and response rate in an online setting where direct access to all potential respondents was limited. To mitigate biases, IP filtering eliminated duplicates, and attention-check items flagged invalid responses (e.g., straight-lining). From 1,107 initial submissions, 865 valid responses were retained based on criteria: completion of secondary education in China and current English-major enrollment. Collection occurred in early 2025, amid ongoing EFL reforms, enhancing timeliness. This timing coincided with increased AI adoption in education, potentially influencing responses on communicative practices (Gao et al., 2025).

Analyses were performed using SPSS 28.0 and AMOS 27.0, following a sequential protocol:

(1) Reliability: Cronbach’s α and item-total correlations assessed internal consistency.

(2) Exploratory Factor Analysis (EFA): On a random sample split (n = 432), principal axis factoring with oblique rotation extracted latent structures, assuming factor correlations.

(3) Confirmatory Factor Analysis (CFA): On the holdout sample (n = 433), model fit was evaluated via χ²/df (<3.0), CFI/TLI (>0.90), RMSEA/SRMR (<0.08) (Hu and Bentler, 1999). Convergent/discriminant validity used AVE and inter-construct correlations.

(4) SEM: A mediation model tested perceptions to attitudes to beliefs to address RQ3, emphasizing attitudes’ mediating role in linking past classroom experiences (perceptions) to future orientations (beliefs). Bootstrapping (5,000 resamples) was used for indirect effects to provide robust, non-parametric estimates, accounting for potential non-normality. Multi-group invariance assessed moderation by gender and starting age, examining whether the mediation paths varied across subgroups.

(5) Comparisons: t-tests/ANOVA explored subgroup differences.

(6) Robustness: Alternative models (e.g., reversed mediation) and Harman’s test checked for common method variance (<50% threshold). Missing data (<1%) were handled via FIML. Additional exploratory analyses examined correlations with learning onset, informed by recent cognition studies (Hsieh and Chuang, 2021). Additionally, to address potential response bias in Likert scales, reverse-scored items were included in subscales, and exploratory multicollinearity checks (VIF < 3.0) confirmed independence. These measures, aligned with applied linguistics standards (Plonsky and Derrick, 2016), mitigate cultural response tendencies in Chinese samples, ensuring mediation model’s validity for cross-cultural replication.

To ensure model robustness, assumptions for SEM were verified: multivariate normality via Mardia’s test (skewness < 3, kurtosis < 10), linearity through scatterplots, and homoscedasticity via Breusch-Pagan tests (all p > 0.05). Potential common method bias beyond Harman’s test was mitigated by procedural remedies, such as varied scale anchors and temporal separation in questionnaire sections. These steps align with best practices in applied linguistics (Plonsky and Derrick, 2016), facilitating replication. Future adaptations could incorporate longitudinal data to test causal inferences in the mediation model.

While SEM allows for testing mediation hypotheses, the cross-sectional design limits causal inferences, as temporal precedence among perceptions, attitudes, and beliefs cannot be established. This risks implying unwarranted causality; thus, results are interpreted as associations, with calls for longitudinal designs in future research (see Limitations in section 6.4) to validate directional paths.

Adhering to the Declaration of Helsinki and institutional protocols, participation was voluntary, anonymous, and withdrawable. No incentives were offered to avoid coercion, and data were stored securely with access limited to researchers. Potential biases (e.g., social desirability) were addressed through neutral item wording and anonymity assurances. Ethical considerations extended to digital data handling, ensuring compliance with data protection standards and enhancing participant confidentiality (Plonsky and Derrick, 2016).

Cronbach’s α ranged from 0.73 to 0.92 across subscales, indicating good internal consistency, except for grammar-focused perceptions (α = 0.599), retained after item analysis confirmed theoretical alignment; sensitivity tests excluding borderline items yielded consistent results. A random sample split enabled EFA (n = 432), extracting three factors (perceptions, attitudes, beliefs) explaining 68.4% variance (KMO = 0.87, Bartlett’s p < 0.001). CFA on the holdout (n = 433) showed strong fit: χ²/df = 2.54, CFI = 0.94, TLI = 0.93, RMSEA = 0.047 (90% CI: 0.042–0.052), SRMR = 0.039. Factor loadings > 0.60 (p < 0.001); AVE > 0.50 for convergent validity; square root of AVE > inter-correlations for discriminant validity (Fornell and Larcker, 1981).

The following are the Cronbach’s α coefficients for each subscale (Table 1).

Finding: These results confirm the instrument’s overall reliability and validity, supporting the study’s objective to quantitatively model teacher cognition components (perceptions, attitudes, beliefs) for addressing all RQs.

Grammar-focused practices predominated (M = 3.74, SD = 0.54) over communication-focused ones [M = 3.23, SD = 0.88; paired t(864) = 18.76, p < 0.001, d = 0.64 (95% CI: 0.57–0.71)]. All grammar items exceeded the midpoint (3.0; one-sample ts > 10.0, ps < 0.001). Communication items like “emphasizing effective communication with occasional grammar” (M = 3.45) suggested partial presence (see Table 2). Exploratory correlations showed early starters perceiving more communicative practices (r = 0.15, p < 0.01).

The following are the means and standard deviations for perceptions of secondary English classroom practices (Table 2).

Finding: Addressing RQ1, pre-service teachers perceive grammar-focused practices as dominant in secondary classrooms, highlighting the influence of exam-oriented systems on CLT cognition, consistent with the objective to explore perception-attitude-belief interrelations.

Attitudes favored communication-focused practices (M = 3.82, SD = 0.69) over grammar-focused [M = 3.29, SD = 0.66; paired t(864) = 22.14, p < 0.001, d = 0.75 (95% CI: 0.68–0.82)]. Highest endorsement for “creating English-use environments” (M = 3.99; see Table 3), echoing pre-service shifts toward CLT amid reforms (Ye et al., 2025). Subgroup analysis indicated females’ stronger communicative attitudes correlated with higher beliefs (r = 0.22, p < 0.001).

The following are the means and standard deviations for attitudes toward classroom practices (Table 3).

Finding: For RQ2, attitudes favor communicative practices despite grammar-dominant perceptions, supporting the objective of examining attitudes as mediators in CLT cognition.

Beliefs leaned toward communication (M = 3.98, SD = 0.63) versus grammar [M = 3.66, SD = 0.60; paired t(864) = 15.92, p < 0.001, d = 0.54 (95% CI: 0.47–0.61)]. Strongest items included “formal grammar for proficiency” (M = 4.04) and “trial-and-error communication improves skills” (M = 4.12; see Table 4), indicating hybrid orientations consistent with evolving EFL cognitions. Early exposure positively correlated with communicative beliefs (r = 0.18, p < 0.001).

The following are the means and standard deviations for beliefs about English language learning (Table 4).

Finding: Responding to RQ3, beliefs lean toward communicative approaches with hybrid elements, aligning with the objective to explore belief formation through attitudes and perceptions.

The mediation model fit well: χ²/df = 2.73, CFI = 0.95, TLI = 0.94, RMSEA = 0.049 (90% CI: 0.044–0.054), SRMR = 0.041. Perceptions predicted attitudes (grammar: β = 0.41, p < 0.001; communication: β = 0.47, p < 0.001). Attitudes predicted beliefs (grammar: β = 0.38, p < 0.001; communication: β = 0.44, p < 0.001). Attitudes mediated perceptions-beliefs, confirming the indirect effect [β = 0.19, bootstrapped 95% CI (0.12, 0.27), p < 0.001; see Figure 1].

The following details the SEM results, including path coefficients, standard errors, t-values, and p-values (Table 5).

Finding: The SEM confirms attitudes’ mediating role in the perceptions-to-beliefs pathway (RQ3), supporting the objective to quantify interrelations in CLT cognition.

Females endorsed communicative attitudes more strongly (M = 3.86) than males [M = 3.71; t(863) = 2.43, p = 0.015, d = 0.22 (95% CI: 0.08–0.36)]. No grammar differences (p > 0.05). Early starters (pre-primary) held stronger communicative beliefs (M = 4.05) than middle/high school starters [M = 3.89; F(2, 862) = 4.21, p = 0.016, η² = 0.01 (95% CI: 0.00–0.02)], with post hoc Tukey confirming significance (p < 0.05). No other subgroup differences were significant. These variations underscore how demographic factors moderate the perceptions-attitudes-beliefs interplay in CLT cognition.

The following summarizes the group comparisons for communicative attitudes and beliefs (Table 6).

Finding: For RQ4, gender and starting age moderate communicative orientations, aligning with the objective to examine variations in cognition interrelations.

Alternative models (reversed mediation: CFI = 0.88, RMSEA = 0.082; direct-only: CFI = 0.89, RMSEA = 0.079) fit poorly, supporting the hypothesized structure. Harman’s test (36% variance) ruled out substantial common method bias. Additional checks, including outlier removal, confirmed results’ stability.

The following compares model fit indices for robustness checks (Table 7).

Further robustness checks included sensitivity analyses excluding the low-reliability grammar-perception subscale (α = 0.599), yielding comparable SEM fit (χ²/df = 2.68, CFI = 0.94) and indirect effect [β = 0.18, 95% CI (0.11, 0.26)]. Subgroup interactions were tested via moderated mediation, revealing no significant gender moderation on the indirect path (index = 0.02, p = 0.45), but a marginal effect for early exposure (index = 0.05, p = 0.08), suggesting experiential factors warrant deeper exploration in larger samples.

Finding: Robustness confirms the model’s validity, supporting all RQs and the objective to reliably quantify mediation and variations in teacher cognition.

Consistent with prior research (Chen et al., 2016; Yan, 2015), participants reported that grammar-focused practices dominated their secondary school classrooms. This dominance is not merely a statistical artifact but reflects a deeper interplay of institutional inertia and cultural expectations, where the Gaokao’s emphasis on accuracy reinforces a “teaching to the test” paradigm that marginalizes communicative fluency (Qi, 2007). Although communicative activities were present to some degree, their secondary role suggests that systemic pressures—such as large class sizes and proficiency disparities—create a self-perpetuating cycle, limiting teachers’ agency to innovate (Chen et al., 2024; Zhu and Shu, 2017). Contextually, this aligns with exam-driven EFL challenges in East Asia, where hierarchical norms further entrench grammar as a proxy for educational rigor (Li, 1998; Gorsuch, 2000; Jin and Yoo, 2019). Theoretically, it illustrates how external constraints shape perceptual schemas early on, potentially foreclosing alternative pedagogies unless disrupted by reform.

Despite their grammar-oriented classroom experiences, pre-service teachers demonstrated stronger support for communicative approaches, especially the creation of environments conducive to English use. This attitudinal preference indicates a form of cognitive dissonance resolution, where exposure to global CLT ideals during university training fosters aspirational views that challenge lived realities, potentially seeding long-term pedagogical change (Savignon and Wang, 2003; Doeur, 2022). Subgroup findings, such as females’ stronger communicative attitudes, may stem from gendered socialization in Chinese education, where relational emphases align with CLT’s interactive ethos (Shan and Li, 2020; Ye et al., 2025). Explanatorily, this highlights attitudes as adaptive mechanisms, bridging the gap between exam-constrained pasts and reform-oriented futures, though implicit-explicit discrepancies underscore unresolved tensions (Sun et al., 2022).

The SEM results extend our understanding of this gap by showing that perceptions of prior classroom practices shaped pre-service teachers’ beliefs indirectly through attitudes. Subgroup variations further reveal moderated interrelations, with females and early starters showing stronger communicative alignments, suggesting experiential and sociocultural influences on cognition dynamics. This mediating effect suggests that beliefs are not direct reflections of past experiences but are filtered through evaluative stances developed during teacher education. Critically extending Borg (2003, 2011) framework, our model refines its dynamic cognition by empirically quantifying attitudes as a pivotal mediator, revealing how they actively reinterpret perceptions in context-specific ways, e.g., moderating exam pressures into hybrid beliefs. However, this extension invites critique: unmodeled moderators like gender and early exposure (as per subgroup effects) may oversimplify the framework in diverse EFL settings, echoing activity theory’s emphasis on sociocultural tools in cognition (Peng, 2024). For instance, while mediation aligns with Borg’s social mediation, it overlooks potential feedback loops where beliefs retroactively influence attitudes, warranting longitudinal refinement (Wang and Zhang, 2024; Kartchava et al., 2020). In China, this underscores cognition as negotiated amid Confucian authority norms, contrasting with less hierarchical contexts (Alharbi, 2022).

Beyond examinations, cultural traditions also shaped participants’ cognitions. The endorsement of statements such as “teachers should provide explicit grammar explanations” reflects the influence of Confucian heritage culture, which positions teachers as authoritative knowledge transmitters (Wei, 2016). This helps explain why grammar-focused beliefs coexist with positive communicative beliefs—a hybrid orientation shaped by both cultural expectations and global pedagogical trends. At the same time, policy reforms such as the New English Curriculum (NEC) promoted by China’s Ministry of Education (Fang and Garland, 2013) may explain the growing emphasis on communicative practices. Pre-service teachers’ strong support for communicative learning suggests that national policies are gradually reshaping pedagogical norms, even if classroom realities have yet to fully align. However, this shift may be uneven due to persistent cultural resistances, such as teacher-centered hierarchies, highlighting the need for policies that integrate local values with CLT adaptations rather than imposing external models.

The findings have several implications:

Teacher education: Programs should explicitly address the belief–practice gap by providing opportunities for pre-service teachers to critically reflect on their prior learning experiences and experiment with communicative methods during teaching practicums. Integrating AI tools can enhance this, as shown in studies on pre-service preparedness for AI-integrated education (Guan et al., 2024).

Curriculum and assessment: Sustainable CLT implementation requires alignment between classroom practices and assessment systems. Without examination reform, teachers’ willingness to adopt communicative methods will remain constrained.

Localized pedagogy: The coexistence of grammar- and communication-focused orientations suggests that CLT in China should evolve as a context-specific approach, integrating global principles with local realities, as seen in similar adaptations elsewhere (Ntirenganya, 2015).

By focusing on pre-service teachers—a group underrepresented in CLT research—this study contributes to the broader teacher cognition literature. It provides empirical evidence that early learning experiences, attitudes, and beliefs interact in complex ways, shaped by both cultural heritage and policy reforms. These insights are particularly relevant to China but offer a cautious lens for other EFL contexts with similar exam and cultural pressures (e.g., Korea, Iran). The quantitative mediation model advances Borg’s framework, though its single-country, cross-sectional, self-report design limits generalizability; future studies should test it longitudinally and cross-culturally to refine CLT theory (Borg, 2015). Critically, while attitudes mediate in this context, sociodemographic moderators (e.g., gender) imply a need for intersectional approaches to disrupt grammar cycles equitably (Shan and Li, 2020; Wang and Zhang, 2024).

Although not the primary focus of this study, AI integration emerges as a promising tool to support CLT in exam-oriented contexts like China. AI can enhance communicative competence by providing personalized feedback, simulating real-world interactions (e.g., via chatbots for role plays), and reducing teacher workload through automated lesson planning and assessment (Guan et al., 2024). For pre-service teachers, AI fosters acceptance of communicative tasks by addressing proficiency gaps and offering equitable access to practice, potentially mitigating the belief-practice gap observed here (Gao et al., 2025; Liu et al., 2024). However, ethical concerns such as data privacy and digital divides must be addressed to ensure inclusive implementation, aligning with localized CLT reforms.

The findings extend the literature on teacher cognition by highlighting the dynamic interplay between perceptions, attitudes, and beliefs during the formative stage of professional development (Borg, 2003, 2011). Specifically aligned with RQ3 and the mediation objective, the results refine our understanding of the belief–practice gap, showing that pre-service teachers’ beliefs are not direct replications of their prior experiences but are refracted through evaluative attitudes shaped by both teacher education and sociocultural contexts. This mediation mechanism contributes to ongoing debates about the contextual nature of teacher cognition and the adaptability of CLT in exam-driven EFL environments, integrating insights from intercultural competence studies (Gong et al., 2022). These conclusions are firmly supported by theories such as Borg’s dynamic cognition framework and activity theory (Peng, 2024), which emphasize mediated, context-sensitive processes; incorporating theories here is necessary to validate interpretations and highlight contributions beyond empirical data.

Several implications emerge for policy and practice: Teacher education programs should create structured opportunities for reflection and practice that challenge pre-service teachers to reconcile exam-oriented traditions with communicative pedagogy. Incorporating AI tools can facilitate this by targeting attitudes as mediators, as evidenced by studies on generative AI acceptance (Gao et al., 2025), where AI enables personalized simulations and feedback to enhance interactive learning. Curriculum and assessment reform is essential. Without alignment between communicative learning goals and high-stakes examinations, teachers’ willingness to implement CLT will remain constrained. Localized adaptation of CLT is necessary. Rather than treating CLT as a universal model, teacher educators and policymakers should acknowledge the hybrid orientation observed in China, where grammar-focused traditions coexist with growing communicative orientations. AI-mediated informal learning offers promising avenues for personalization (Liu et al., 2024). By leveraging attitudes as mediators and addressing subgroup variations (RQ4), policies could prioritize AI-driven simulations in curricula, fostering equitable shifts in cognition across EFL regions. supporting equitable shifts in cognition. By leveraging attitudes as mediators, policies could prioritize AI-driven simulations in curricula, fostering equitable shifts in cognition across EFL regions. This approach not only bridges gaps in China but offers scalable models for global adaptation, emphasizing empirical testing in comparative studies (Song et al., 2025).

Although the study is situated in China, its findings resonate with other EFL contexts—such as Korea, Japan, Iran, and Saudi Arabia—where cultural traditions and exam pressures shape teachers’ pedagogical orientations. By centering pre-service teachers, this research highlights how future generations of educators may bridge the tension between global pedagogical trends and local educational realities. The incorporation of AI insights, particularly in mediating attitudes, extends relevance to digital-era EFL, addressing equitable education challenges.

The study is not without limitations. First, reliance on self-reported survey data raises concerns about social desirability bias. Second, the focus on a single region in China may limit the generalizability of findings to other regions. Third, the cross-sectional design cannot capture the longitudinal evolution of teacher cognition.

Future research should therefore adopt mixed-method approaches, combining surveys with interviews and classroom observations to triangulate findings. Longitudinal studies tracking pre-service teachers into their in-service years would provide valuable insights into how attitudes and beliefs translate into practice over time. Additionally, comparative studies across multiple EFL contexts could deepen understanding of how cultural, institutional, and policy factors mediate the global-local dynamics of CLT, building on this study’s mediation and variation findings (RQ3 and RQ4). Exploring AI’s impact on cognition, such as in GenAI for teacher identity and corrective feedback, is a promising direction (Song et al., 2025; Kartchava et al., 2020).

To overcome these limitations, policymakers could fund hybrid training programs integrating AI simulations with practicum experiences, targeting attitude mediation to accelerate CLT shifts. Comparative research in diverse EFL regions (e.g., Vietnam or Iran) could validate the model’s cross-cultural applicability, refining global standards for teacher cognition amid technological advancements (Borg, 2015; Song et al., 2025).