VLM-fusion: an intelligent diagnostic system for rural teacher professional development integrating vision-language models and adaptive learning path optimization

Abstract

Rural teachers face persistent professional development challenges stemming from geographic isolation, limited training resources, and scarce peer collaboration opportunities. This paper presents an intelligent diagnostic system that integrates vision-language model (VLM) capabilities with adaptive learning path optimization to deliver personalized, context-aware development support for rural educators. We first construct a four-dimensional competency model encompassing pedagogical content knowledge, practical teaching skills, information and digital literacy, and rural contextual adaptability, with weights determined via the Analytic Hierarchy Process. A multimodal feature fusion diagnostic model employing cross-modal attention mechanisms jointly reasons over classroom video, teaching design texts, artifacts, and supplementary data streams, while a gated fusion strategy ensures robustness under missing-modality conditions common in rural settings. An Actor-Critic reinforcement learning algorithm generates adaptive learning paths by optimizing a composite reward function that balances competency improvement, resource accessibility, and temporal feasibility constraints. Experiments on a dataset of 386 rural teachers from 47 schools across five provinces, with competency ground truth established through structured expert panel ratings, suggest that the proposed VLM-Fusion model attains an F1-score of 84.9% (mean ± 0.6% across five runs) and a Spearman rank correlation of 0.918 in competency diagnosis, comparing favorably against unimodal and conventional fusion baselines. A 12-week quasi-experimental field study (rather than a long-term randomized controlled design) with 96 teachers further indicates that the adaptive path recommendation is associated with a mean competency improvement of 23.7%, exceeding expert-curated (14.2%) and self-directed (8.9%) conditions, while narrowing within-group competency variance over time. We wish to flag upfront that these improvement figures were derived using the same multimodal diagnostic pipeline that generated the personalized recommendations, not an entirely independent assessment procedure; readers should therefore treat the numerical magnitude with appropriate caution, as we discuss at length in the Limitations.

Introduction

The professional growth of teachers in rural and remote areas has long been constrained by a constellation of structural barriers that differ markedly from those encountered by their urban counterparts. Geographic isolation, scarce access to high-quality training resources, and limited peer collaboration opportunities collectively hinder the continuous competence development of rural educators¹. In many developing regions, rural teachers account for a substantial proportion of the teaching workforce, yet the investment in their in-service training remains disproportionately low, resulting in widening pedagogical skill gaps between urban and rural schools². This imbalance not only undermines educational equity but also threatens the quality of foundational learning experiences available to millions of students in underserved communities. Against this backdrop, the accelerating digital transformation of education—propelled by breakthroughs in artificial intelligence and multimodal computing—has opened new avenues for rethinking how professional development support can be designed, delivered, and personalized at scale.

Within this wave of technological advancement, vision-language models (VLMs) have emerged as a particularly promising class of multimodal architectures. By jointly learning from visual and textual data, VLMs such as CLIP, BLIP-2, and their successors demonstrate a remarkable capacity to interpret complex, context-rich information spanning images, documents, and instructional scenarios³. In the educational domain, researchers have begun to explore VLM-based applications for automated classroom observation analysis, teaching behavior recognition, and multimodal learning analytics⁴. These efforts signal a departure from traditional text-only diagnostic approaches, enabling systems that can process heterogeneous data sources—including lesson videos, handwritten artifacts, and visual teaching materials—to generate richer, more holistic assessments of teaching practice⁵. Meanwhile, intelligent diagnostic systems grounded in deep learning and knowledge graph techniques have attracted growing scholarly attention. Such systems seek to identify individual competency profiles, detect latent knowledge deficiencies, and deliver targeted feedback through data-driven inference mechanisms⁶. Recent work has demonstrated that coupling diagnostic outputs with adaptive learning path recommendation algorithms can substantially improve the efficiency and relevance of professional training programs⁷.

Despite these advances, a careful examination of the existing literature reveals several notable shortcomings when it comes to addressing the unique needs of rural teachers. Most intelligent diagnostic frameworks reported to date have been developed and validated within urban or well-resourced institutional contexts, with little consideration for the infrastructural and socio-cultural realities that characterize rural teaching environments⁸. The data modalities employed in prior studies tend to be relatively homogeneous—predominantly survey-based or log-based—overlooking the rich but unstructured visual and contextual information that pervades everyday rural classroom practice⁹. Furthermore, current adaptive learning path planning methods largely rely on static prerequisite graphs or collaborative filtering, which struggle to accommodate the highly heterogeneous starting points and fragmented learning schedules typical of rural teacher populations¹⁰. The integration of multimodal perception capabilities offered by VLMs into the diagnostic loop remains underexplored, and no existing study has systematically addressed how such integration could enhance both the granularity of competency diagnosis and the responsiveness of subsequent learning recommendations for rural educators specifically.

It is precisely this gap that motivates the present research. We argue that a meaningful leap in rural teacher professional development support requires moving beyond unimodal, context-agnostic diagnostic paradigms toward a framework that can perceive, interpret, and act upon the diverse informational signals embedded in authentic rural teaching contexts. By harnessing the cross-modal reasoning strengths of contemporary VLMs and coupling them with adaptive path optimization algorithms sensitive to individual progress trajectories, it becomes feasible to construct an intelligent system capable of delivering genuinely personalized, context-aware development guidance. Such an approach holds both theoretical significance—advancing our understanding of how multimodal AI can serve equity-oriented educational goals—and practical value, offering a scalable mechanism to narrow the professional development divide between urban and rural teachers¹¹.

This paper presents the design, implementation, and evaluation of an intelligent diagnostic system for rural teacher professional development that fuses vision-language model capabilities with adaptive learning path optimization. The research unfolds along three interrelated threads. First, we construct a multimodal data representation framework that captures classroom visual features, textual teaching reflections, and structured competency indicators, enabling comprehensive teacher profiling. Second, we propose a VLM-driven diagnostic module that performs fine-grained competency assessment by jointly reasoning over visual and linguistic evidence extracted from authentic teaching scenarios. Third, we develop an adaptive learning path optimization algorithm that dynamically adjusts recommended training sequences based on real-time diagnostic feedback, individual learning pace, and resource availability constraints characteristic of rural settings. The technical route progresses from multimodal data preprocessing through model fine-tuning and diagnostic inference to path generation and iterative refinement, forming a closed-loop system architecture¹². The main contributions of this work center on three aspects: an early integrated attempt to apply VLMs to rural teacher competency diagnosis, a context-sensitive adaptive path planning mechanism that accounts for rural-specific constraints, and an end-to-end framework that bridges the gap between multimodal perception and actionable professional development recommendations. We note that, while prior studies have explored multimodal analytics and adaptive learning independently, few have combined both within a single closed-loop system tailored to rural educators; existing frameworks have generally not addressed the data incompleteness and resource scarcity that characterize rural teaching environments^8,9. Through this integrated effort, we aim to contribute both a methodological advance in AI-supported educational diagnostics and an empirically grounded solution to a pressing equity challenge in teacher development.

Theoretical foundation and technical framework

Construction of the rural teacher professional development competency model

Any robust diagnostic system must rest on a clearly articulated competency framework that reflects both general pedagogical principles and the particularities of the target population. Drawing on Shulman’s foundational taxonomy of teacher knowledge and subsequent extensions by scholars working in resource-constrained educational settings, we recognize that rural teachers operate under conditions that demand a broader and more adaptive skill repertoire than conventional models typically capture¹³. Standard professional development frameworks—while valuable—tend to privilege content knowledge and generic instructional strategies, often neglecting the contextual competencies that prove decisive in rural classrooms, such as multi-grade teaching adaptability, community engagement capacity, and the ability to improvise with limited technological infrastructure¹⁴.

To address this gap, we constructed a multi-dimensional competency indicator system through an iterative process involving three steps. First, an initial pool of 32 candidate indicators was generated from a systematic review of teacher competency frameworks in the literature^13,14. Second, seven experts participated in iterative consultation rounds to refine and finalize the indicator set. These experts comprised three university-based teacher educators with over ten years of experience in rural education research, two provincial-level teaching and research officers responsible for rural teacher training policy, and two senior rural school principals recognized for instructional leadership. Through two rounds of Delphi-style consultation, the panel reached consensus on 8 secondary indicators organized under 4 primary dimensions, discarding 24 items judged redundant, ambiguous, or insufficiently observable in rural contexts. Third, the resulting framework was piloted with 120 practicing rural teachers across three provinces, and exploratory factor analysis (EFA) confirmed a four-factor structure consistent with the theoretical design (Kaiser–Meyer–Olkin measure = 0.83; Bartlett’s test of sphericity p < 0.001), with all items loading above 0.45 on their intended factors. Because the four primary dimensions are expected, on theoretical grounds, to covary rather than to stand fully independent of one another, we extracted factors via principal axis factoring (PAF) and applied a promax oblique rotation with κ = 4; the inter-factor correlations fell within the range 0.28–0.41, which reinforces the appropriateness of an oblique—rather than orthogonal—rotation scheme. We acknowledge that a pilot sample of 120 teachers across 8 secondary indicators yields a subject-to-item ratio of 15:1, which is adequate but not generous by EFA conventions, and we flag this as a reason for caution when generalizing the factor structure. Cronbach’s alpha coefficients for the four dimensions ranged from 0.78 to 0.86, indicating acceptable internal consistency reliability. The resulting framework encompasses four primary dimensions: pedagogical content knowledge (), practical teaching skills (), information and digital literacy (), and rural contextual adaptability (). Each primary dimension is further decomposed into secondary indicators that operationalize specific observable competencies. Table 1 summarizes the complete indicator hierarchy along with the assigned weights determined through subsequent analysis.

Table 1.

Rural teacher professional development competency indicator system and weights.

Primary dimension	Weight	Secondary indicator	Weight
Pedagogical content knowledge ()	0.32	Curriculum comprehension and instructional design	0.18
Pedagogical content knowledge ()	0.32	Subject knowledge depth and cross-disciplinary awareness	0.14
Practical teaching skills ()	0.28	Classroom management and multi-grade instruction	0.16
Practical teaching skills ()	0.28	Student assessment and differentiated feedback	0.12
Information and digital literacy ()	0.22	Educational technology integration capability	0.12
Information and Digital Literacy ()	0.22	Digital resource retrieval and adaptive tool adoption	0.10
Rural contextual adaptability ()	0.18	Community-school collaboration and local curriculum development	0.10
Rural contextual adaptability ()	0.18	Resilience under resource constraints and self-directed growth	0.08

The weight assignment followed the Analytic Hierarchy Process (AHP), a well-established multi-criteria decision method particularly suited to problems where both quantitative data and qualitative expert judgment must be reconciled¹⁵. Each of the seven experts independently completed pairwise comparison matrices at both the primary-dimension and secondary-indicator levels using Saaty’s standard 1–9 scale. We aggregated individual judgments via the geometric mean method and computed priority vectors through eigenvalue decomposition. The complete pairwise comparison matrix for the four primary dimensions is presented in Appendix Table A1 (provided in the supplementary materials uploaded to the submission system). Inter-expert agreement was assessed using Kendall’s coefficient of concordance, yielding W = 0.81 (p < 0.01), indicating strong consensus among the panel members. We should state plainly, to preempt any concern about circular reasoning: the seven-expert panel that determined the AHP weights and the three-person rating panels that produced the ground-truth competency annotations (Sect. 4.1) were constituted from disjoint expert pools, with no individual serving on both. This separation was deliberate. It prevents the situation in which the same expert’s prior weighting scheme could, through label generation, propagate back into the training loss as a self-fulfilling optimization target. For a given comparison matrix  A of order  n, the principal weight vector  w satisfies:

1

where denotes the largest eigenvalue of  A¹⁶. To verify judgment consistency, we computed the consistency ratio following Saaty’s standard procedure:

2

Here  RI represents the random consistency index corresponding to matrix order  n; all matrices in our analysis yielded  CR<0.10, confirming acceptable coherence among expert assessments¹⁷. With the validated weight structure in place, the composite competency score  S for an individual teacher is then computed as a weighted aggregation across all secondary indicators:

3

where represents the normalized score on the  j-th secondary indicator within the  i-th primary dimension, and is the number of secondary indicators under dimension  i. This formulation yields a scalar proficiency profile that serves as the foundational input to the downstream intelligent diagnostic module. Importantly, the modular design of this scoring model allows future refinement—should new competency facets emerge from evolving rural education policies, additional indicators can be incorporated without restructuring the overall computational logic¹⁸.

Application mechanism of vision-language models in education

Vision-language models represent a class of multimodal architectures that learn joint representations across visual and textual modalities, enabling machines to reason about images and language in a unified semantic space. The conceptual breakthrough underlying most contemporary VLMs is contrastive pre-training on large-scale image-text pairs, an idea crystallized by the CLIP framework³ and extended at even larger scale through noisy-supervision variants such as ALIGN¹⁹, both of which demonstrated that aligning visual encoders and text encoders through a shared embedding objective could yield surprisingly transferable representations without task-specific fine-tuning. BLIP and its successor BLIP-2 refined this paradigm further by introducing generative objectives alongside contrastive ones, allowing the model not only to match images with descriptions but also to produce contextually grounded textual outputs conditioned on visual inputs²⁰. What makes these architectures particularly compelling for educational applications is their capacity to process the kind of heterogeneous, loosely structured data that pervades real teaching environments—classroom photographs, handwritten lesson plans, slide presentations, video keyframes—without requiring rigid, predefined feature extraction pipelines.

Formally, the core alignment mechanism in contrastive VLMs can be expressed as maximizing the cosine similarity between matched image-text pairs while minimizing it for unmatched ones. Given a batch of  N image-text pairs, the contrastive loss for the image-to-text direction takes the standard form:

4

where and denote the normalized embeddings of the  i-th image and its paired text, represents cosine similarity, and is a learnable temperature parameter that controls the sharpness of the distribution²¹. A symmetric loss operates in the reverse direction, and the total training objective combines both:

5

This dual-direction alignment ensures that the learned embedding space is coherent from both perceptual and linguistic perspectives, a property that proves essential when downstream tasks demand cross-modal inference²².

In educational contexts, these capabilities translate into three concrete application pathways that are directly relevant to the diagnostic system proposed in this paper. The first concerns classroom behavior recognition. Recent studies have shown that fine-tuned VLMs can classify teacher-student interaction patterns from video frames with accuracy rates exceeding those of conventional computer vision pipelines, largely because the textual branch supplies semantic priors about pedagogical behaviors that pure visual models lack²³. The second pathway involves instructional resource comprehension—extracting pedagogical intent and content structure from visually rich materials such as annotated worksheets, concept maps, and multimedia courseware. Here, the generative capacity of BLIP-type models proves indispensable, as it enables the system to produce structured textual descriptions of visual teaching artifacts that can then feed into diagnostic reasoning modules²⁴. The third, and perhaps most challenging, pathway pertains to holistic teaching competency assessment. By aggregating cross-modal evidence from multiple data streams—observed classroom behaviors, resource design quality, and reflective textual narratives—a VLM-based system can construct a richer evaluative portrait than any single-modality approach would permit²⁵. We note, however, that realizing this potential in rural settings demands careful domain adaptation, since pre-trained VLMs carry distributional biases toward well-resourced, urban-centric educational imagery, a challenge we address explicitly in the system design described in subsequent Sect²⁶.

Adaptive learning path optimization theory

The notion that instruction should adapt to the evolving state of each learner is far from new—its intellectual roots stretch back to Crowder’s branching programmed instruction in the 1960s—but the computational machinery to realize this vision at scale has only matured in recent years. Knowledge tracing, the task of modeling a learner’s latent mastery over time, has undergone successive paradigm shifts: from early Bayesian Knowledge Tracing models built on hidden Markov assumptions, through item response theory variants, to contemporary deep knowledge tracing architectures that harness recurrent and attention-based neural networks to capture complex temporal learning dynamics²⁷. These advances have been instrumental in providing the state estimation backbone that any adaptive path planner requires. Yet accurate diagnosis of what a learner knows is only half the problem. The harder question—and the one more directly pertinent to our system—is how to translate that evolving diagnosis into a sequence of learning activities that optimally advances the learner toward specified competency targets.

Reinforcement learning (RL) offers an elegant formalism for this sequential decision problem. Rather than prescribing a fixed curriculum, an RL-based path planner treats each recommendation as an action taken in response to the learner’s current state, receiving reward signals tied to observed competency gains²⁸. Concurrently, graph neural networks (GNNs) have proven effective at encoding the relational structure among knowledge concepts and training modules, capturing prerequisite dependencies and semantic affinities that flat vector representations inevitably miss²⁹. Several recent works have demonstrated that coupling GNN-derived knowledge graph embeddings with RL-based policy learning yields path recommendations superior to those generated by either technique in isolation³⁰.

We ground our adaptive path optimization module in the Markov Decision Process (MDP) formalism, a standard framework for sequential decision-making under uncertainty. The learning path planning problem is formulated as a tuple , where denotes the set of learner competency states derived from the diagnostic module described in Sect. 2.1, represents the set of recommendable learning activities,  P is the state transition function,  R is the reward function, and is the discount factor³¹. The optimization objective is to find a policy that maximizes the expected cumulative discounted reward:

6

where and represent the state and action at step  k, and  T is the planning horizon. The reward function  R is designed to reflect competency improvement as measured by changes in the composite score  S defined in Eq. (3), with an additional penalty term that accounts for resource accessibility constraints specific to rural settings:

7

Here, captures the marginal competency gain following action , quantifies the resource cost associated with the recommended activity (e.g., bandwidth requirements, material availability), and , are weighting coefficients that balance learning effectiveness against practical feasibility³². One classical approach to solving this MDP is the deep Q-network (DQN), which approximates the optimal action-value function through iterative gradient-based updates on experience replay buffers³³. For conceptual clarity, we present the Bellman optimality equation that underlies value-based RL methods:

8

However, DQN-based methods face practical limitations in our setting: the learning resource action space is large and partially continuous (combining discrete resource selection with continuous scheduling parameters), and DQN’s deterministic greedy policy offers limited exploration. We therefore implement the actual path optimization using an Actor-Critic architecture trained with Proximal Policy Optimization (PPO), as detailed in Sect. 3.3. The Bellman equation above remains relevant as the theoretical foundation; the Actor-Critic framework can be understood as an alternative solution strategy for the same underlying MDP that replaces explicit Q-function maximization with policy gradient updates guided by a learned value baseline⁴⁵. The integration of GNN-encoded prerequisite structures into the state representation ensures that the policy respects the topological ordering of competency development—an aspect often overlooked by generic RL approaches but crucial for coherent professional development sequencing in practice.

Intelligent diagnostic system design and algorithm implementation

Overall system architecture design

Translating the theoretical constructs outlined in Section II into a functioning diagnostic system demands an architecture that can accommodate heterogeneous data inputs, perform cross-modal reasoning, and deliver actionable developmental guidance—all while remaining responsive enough for iterative use in bandwidth-constrained rural environments. After considering several candidate designs, we settled on a four-layer architecture that separates concerns cleanly while maintaining tight data flow integration across layers. Figure 1 illustrates the complete system structure, depicting the functional modules, their internal components, and the directional flow of information from raw data ingestion through to personalized path delivery.

Fig. 1.

Overall architecture of the VLM-based intelligent diagnostic system for rural teacher professional development (note: the path recommendation component employs an Actor-Critic/PPO architecture as described in Sect.  3.3).

The foundation of the system is the Data Acquisition Layer, which serves as the primary interface between the real-world teaching environment and the computational pipeline. This layer aggregates multimodal inputs from diverse sources: classroom observation videos captured via low-cost recording devices, digitized lesson plans and teaching reflections uploaded through a lightweight mobile application, structured self-assessment questionnaires aligned with the competency indicators defined in Table 1, and contextual metadata describing the teacher’s school resources and geographic conditions. A key design decision here was to support asynchronous, offline-capable data collection—a practical necessity given that many rural schools experience intermittent network connectivity³⁴. Raw inputs undergo preliminary validation and format standardization before being forwarded to the subsequent processing stage.

Sitting above this ingestion layer, the Feature Extraction Layer transforms heterogeneous raw data into semantically meaningful representations amenable to diagnostic reasoning. Visual data streams—video keyframes, photographs of classroom artifacts, scanned handwritten materials—pass through the visual encoder of our fine-tuned VLM, while textual inputs are processed by the corresponding language branch. The contrastive alignment mechanism described in Eqs. (4) and (5) ensures that extracted visual and textual features reside in a shared embedding space, enabling direct cross-modal comparison and fusion. For structured numerical inputs such as questionnaire scores and institutional metadata, we employ a separate encoding submodule based on a lightweight multi-layer perceptron, whose outputs are projected into the same representational space via a learned linear transformation³⁵. The fused multimodal feature vectors then constitute the state representation that feeds both the diagnostic engine and the path planner.

The Intelligent Diagnostic Layer constitutes the analytical core of the system. Receiving the fused feature representations, this module performs two complementary functions: dimension-level competency scoring and holistic profile classification. Dimension-level scoring maps each teacher’s multimodal evidence onto the four primary competency dimensions ( through ) using the weighted aggregation model in Eq. (3), producing fine-grained assessments that pinpoint specific developmental needs. Holistic profile classification, meanwhile, assigns each teacher to one of several developmental archetypes—such as “content-strong but digitally underequipped” or “practically skilled but lacking reflective depth”—through a clustering mechanism trained on historical diagnostic data³⁶. Together, these outputs provide both the quantitative precision and the qualitative interpretability necessary for meaningful professional development guidance.

Finally, the Path Recommendation Layer consumes the diagnostic outputs and generates personalized, sequenced learning activity plans. Grounded in the MDP framework formalized in Eq. (6) through (8), this layer employs an Actor-Critic policy network augmented with GNN-encoded knowledge graph structure to select and order training resources, micro-courses, peer collaboration opportunities, and practice tasks that align with each teacher’s diagnosed deficiencies and available resource conditions. Critically, the path recommendation module operates in a closed-loop fashion: as teachers engage with recommended activities and new multimodal evidence accumulates, the diagnostic layer updates its competency estimates, which in turn trigger path recalibration. This iterative feedback cycle—from data acquisition through diagnosis to recommendation and back—ensures that the system adapts continuously rather than offering a single static prescription, a property we regard as essential for sustaining engagement among rural teachers whose professional development trajectories are rarely linear or predictable³⁷.

Multimodal feature fusion diagnostic model

The diagnostic accuracy of our system hinges on its ability to construct a unified representation from data sources that differ not merely in format but in fundamental semantic granularity. A classroom video keyframe captures spatial-temporal behavioral patterns; a teaching design document encodes pedagogical intentionality through linguistic structure; student feedback data reflects learning outcomes in quantified or semi-structured form. Naively concatenating features extracted independently from these modalities—an approach still common in many educational analytics systems—discards precisely the inter-modal correlations that carry the most diagnostic signal³⁸. Our design instead pursues deep fusion through a cross-modal attention mechanism that allows each modality to selectively attend to informative elements in the others before any diagnostic inference occurs.

We begin by specifying how each modality is independently encoded. Table 2 summarizes the data types, feature extraction methods, output dimensionality, representational characteristics, and fusion strategy for each modality stream processed by the system.

Table 2.

Multimodal feature dimensions and fusion strategies.

Data modality	Source description	Extraction method	Feature dimension	Fusion strategy
Classroom video	Keyframes sampled at 1 fps from recorded lessons	VLM visual encoder (ViT-L/14)		Cross-modal attention
Teaching design text	Lesson plans, instructional reflections	VLM text encoder (Transformer-based)		Cross-modal attention
Student feedback	Structured questionnaires and rating scales	Multi-layer perceptron encoder		Linear projection + concatenation
Classroom artifacts	Scanned worksheets, board writing photographs	VLM visual encoder (ViT-L/14)		Cross-modal attention
Contextual metadata	School resource indices, class size, connectivity	Tabular embedding network		Gated concatenation
Self-assessment scores	Teacher competency self-evaluation responses	Normalized scalar encoding		Weighted summation

As Table 2 shows, the three primary modalities—video, text, and artifacts—share the same embedding dimensionality by virtue of originating from the same VLM backbone, which considerably simplifies the subsequent fusion mechanism. Supplementary modalities (feedback, metadata, self-assessment) are projected into compatible spaces through learned transformations.

The core of our fusion approach is a cross-modal attention module inspired by recent advances in multimodal Transformer architectures, though adapted here for the specific demands of educational diagnostic reasoning³⁹. Let , , and represent the sequence-level feature matrices for video, text, and artifact modalities respectively, where , , denote token counts and  d is the shared embedding dimension. For any pair of modalities—taking video-to-text attention as an illustrative case—the cross-modal attended representation is computed as:

9

where are learnable projection matrices for queries, keys, and values respectively, and denotes the projected dimension⁴⁰. This formulation enables video features to be enriched by attending to semantically relevant segments of the teaching design text—a teacher’s stated pedagogical intention can, for instance, contextualize an otherwise ambiguous classroom behavior pattern. Symmetric attention operations produce text-attended-by-video and artifact-attended-by-text representations, yielding a family of cross-attended feature matrices.

These enriched modality-specific representations must then be consolidated into a single diagnostic feature vector. We accomplish this through a gated fusion mechanism that adaptively determines each modality’s contribution based on input quality—a particularly important consideration given that rural data collection conditions often produce incomplete or degraded modality streams. The fused representation  h is obtained as:

10

11

In Eqs. (10) and (11), is the cross-attended feature for modality  m, denotes the sigmoid activation, is the learned gating vector, represents element-wise multiplication, and  [f;c;s] is the concatenation of supplementary features from student feedback, contextual metadata, and self-assessment, projected through weight matrix ⁴¹. The gating mechanism ensures graceful degradation: when a video stream is missing or corrupted, the corresponding gate values approach zero, allowing textual and artifactual evidence to compensate without catastrophic information loss.

The final diagnostic inference maps the fused representation  h to competency scores across the four dimensions defined in Sect.  2.1. We formulate this as a multi-task regression problem where dimension-specific prediction heads share the fused backbone but maintain independent output layers:

12

Here, is the predicted normalized score for the  j-th secondary indicator under the  i-th primary dimension, and the sigmoid output ensures scores remain within . The training objective minimizes the mean squared error between predicted scores and ground-truth annotations collected from expert evaluators:

13

We acknowledge that AHP weights reflect expert judgments about the relative importance of competency dimensions, and using them directly as loss-weighting coefficients assumes that prediction errors in more important dimensions should receive proportionally greater optimization pressure. This assumption is not self-evident; an alternative view is that all dimensions deserve equal optimization effort regardless of their assessed importance. To examine this, we conducted a comparison with uniform loss weighting (all weights set to 1/8), which yielded an F1-score of 82.6% versus 84.9% for the AHP-weighted variant, suggesting that the importance-proportional weighting does confer a modest but consistent benefit. For transparency on stability rather than mean values alone, we note that across the five independent runs the AHP-weighted model outperformed its uniform-weighted counterpart in all five cases, with the per-run F1 gap falling within the narrow band of 1.9–2.7% points (mean 2.3, SD 0.3). The advantage is therefore not a seed-dependent artifact; it reproduces reliably. The rationale is that dimensions weighted more heavily by experts tend to be those where diagnostic errors carry larger downstream consequences for path recommendation quality. In Eq. (13), the AHP-derived weights and from Eqs. (1) and (2) serve double duty here as loss-weighting coefficients, ensuring that the model’s optimization effort is proportionally allocated to competency dimensions according to their assessed importance⁴². The predicted scores can be directly substituted into Eq. (3) to produce the composite competency score  S, which then feeds into the adaptive path optimization module described in Sect. 2.3, thereby closing the link between multimodal perception and actionable developmental planning.

Adaptive learning path generation algorithm

With diagnostic outputs providing a granular portrait of each teacher’s competency profile, the remaining challenge is to translate those assessments into a coherent, temporally ordered sequence of learning activities that respects both pedagogical logic and the practical realities of rural professional life. The first step toward this goal involves constructing a learning resource association network—essentially a knowledge graph whose nodes represent individual training modules, micro-courses, practice tasks, and peer mentoring opportunities, and whose edges encode prerequisite dependencies, thematic affinity, and difficulty progression relationships. We build this graph by combining expert-curated curriculum structures with data-driven edge weight estimation based on historical learner transition patterns⁴³. Each node in the resource graph is characterized by an attribute vector that includes its target competency dimension, estimated completion time, bandwidth requirement, and offline availability status—attributes whose importance becomes clear when we consider that a rural teacher may have only sporadic evening hours and unreliable internet access to devote to professional learning.

To propagate structural information through this graph and produce resource embeddings that capture relational context, we apply a two-layer graph attention network (GAT) over . The embedding of resource node after message passing is given by:

14

where denotes the neighborhood of node  i, is a shared weight matrix, and are attention coefficients computed via a single-layer feedforward network followed by softmax normalization over the neighborhood, consistent with the standard GAT formulation⁴⁴. These graph-aware embeddings serve as the action representations within the reinforcement learning framework.

For reproducibility, Table 3 lists the key implementation details and hyperparameters used across all model components.

Table 3.

Implementation details and hyperparameters.

Component	Parameter	Value
VLM Backbone	Architecture	BLIP-2 (ViT-L/14 + OPT-2.7B)
	Pre-training checkpoint	Salesforce/blip2-opt-2.7b (HuggingFace)
Fine-tuning	Optimizer	AdamW
	Learning rate	2e-5 (cosine decay)
	Batch size	16
	Epochs	30 (early stopping, patience = 5)
	Modality dropout rate	0.15 per modality per sample
Video processing	Sampling strategy	1 fps keyframe, max 900 frames/clip
Text processing	Token limit	512 tokens per document
Missing modality simulation	Training procedure	Random single-modality dropout
PPO (path optimization)	Clipping threshold	0.2
	Actor learning rate	3e-4
	Critic learning rate	1e-3
	Epochs per update	10
	Discount factor	0.99
Knowledge graph	Nodes	247 learning resources
	Edges	1,034 (prerequisite: 412; thematic: 389; difficulty: 233)
	Construction	Expert-curated topology + data-driven edge weights from 2,847 historical transition records
Data split	Procedure	Teacher-level stratified split: 80% train, 20% test
Hardware	Training	4 x NVIDIA A100 (40 GB), ~ 18 h total
	Inference	Single A100, ~ 2.3 s per teacher profile
Baseline tuning	Protocol	Grid search on validation set (same split)
Stopping criteria	Diagnostic model	Validation F1 no improvement for 5 epochs

We adopt an Actor-Critic architecture to solve the path optimization problem, a choice motivated by its superior stability over pure policy gradient methods in high-dimensional action spaces and its ability to produce stochastic policies that naturally incorporate exploration—a desirable property when the optimal learning path for a given teacher profile is uncertain⁴⁵. Figure 2 depicts the complete algorithmic workflow, from diagnostic state construction through actor-critic interaction to path output and feedback-driven policy update.

Fig. 2.

Flowchart of the adaptive learning path generation algorithm based on deep reinforcement learning.

As Fig. 2 illustrates, the algorithm operates in an iterative loop. At each decision step  k, the actor network receives the current state —a concatenation of the teacher’s fused diagnostic feature vector  h from Eq. (11), the graph embedding of the most recently completed resource, and a temporal context vector encoding elapsed training time and remaining schedule capacity. The actor outputs a probability distribution over candidate next actions (i.e., learning resources), from which action is sampled:

15

Simultaneously, the critic network estimates the expected cumulative reward from state onward, providing the baseline against which actual returns are measured. The advantage function, which drives the policy gradient, takes the standard temporal-difference form:

16

A critical contribution of our work lies in the design of the reward function , which extends the basic formulation in Eq. (7) to account for the multifaceted constraints that shape rural teachers’ learning trajectories. We propose a composite objective that balances competency improvement, resource accessibility, and temporal feasibility:

17

Here, beyond the competency gain and resource cost defined previously, represents the estimated time required for activity , is the teacher’s declared available time at step  k, and the penalty term ensures that the algorithm does not recommend activities that exceed realistic time budgets—a constraint that we found essential during pilot testing, as rural teachers frequently reported abandoning training programs whose individual modules demanded uninterrupted time blocks incompatible with their schedules⁴⁶. The final term rewards coverage diversity, measured as the marginal increase in the number of distinct competency sub-dimensions addressed by the recommended path so far, discouraging repetitive recommendations that fixate on a single weakness. The coefficients , , , and are tunable hyperparameters determined through grid search on a validation set.

The actor parameters are updated via the proximal policy optimization (PPO) clipped objective to ensure training stability:

18

where is the importance sampling ratio and ε is the clipping threshold, following the standard PPO formulation. The critic is updated by minimizing the squared temporal-difference error:

19

Both networks are trained jointly through alternating gradient updates on trajectory batches collected from simulated teacher learning episodes⁴⁷. During deployment, the trained actor generates path recommendations in real time; as new diagnostic evidence accumulates after each completed activity, the state is refreshed and the next recommendation reflects the teacher’s updated competency profile, ensuring genuine adaptivity rather than static path assignment.

Experimental verification and result analysis

Experimental design and dataset construction

Validating the proposed system requires a dataset that faithfully reflects the multimodal complexity of rural teaching practice—something that, to our knowledge, no publicly available benchmark currently provides. We therefore undertook a dedicated data collection effort spanning 18 months across 47 rural primary and secondary schools located in five provinces with varying levels of economic development and geographic accessibility. Participating teachers ( N=386) were recruited through local education bureaus and consented to contribute classroom recordings, lesson planning documents, written teaching reflections, student feedback questionnaires, and institutional metadata. Each teacher contributed data across two to four observation cycles spaced approximately one semester apart, yielding a longitudinal dimension that proved essential for validating the system’s adaptive path recommendations over time.

Table 4 presents the statistical profile of the resulting dataset, broken down by modality type.

Table 4.

Dataset statistical information.

Data modality	Total samples	Avg. per teacher	Format description
Classroom video segments	2,314	5.99	15-min clips, 720p, 1 fps keyframe extraction
Lesson plan documents	1,847	4.78	Digitized text, avg. 1,200 words per document
Teaching reflections	1,562	4.05	Free-text narratives, avg. 650 words
Student feedback forms	3,890	10.08	5-point Likert scale, 12 items per form
Contextual metadata records	386	1.00	School resource index, connectivity status, class size

As Table 4 shows, video and textual data constitute the richest modality streams, while student feedback provides the densest per-teacher sampling. All data underwent a standardized preprocessing pipeline: videos were segmented and keyframes extracted at one frame per second; textual documents were tokenized and cleaned of personally identifiable information; Likert-scale responses were normalized to ; and metadata fields were encoded following the scheme described in Sect. 3.1. Competency ground-truth labels were generated through a rigorous expert annotation process in which panels of three experienced teacher educators independently rated each teacher on all secondary indicators from Table 1, with final scores computed as the panel mean after discarding outliers exceeding two standard deviations from consensus. Inter-rater reliability was assessed using the intraclass correlation coefficient (ICC, two-way random, absolute agreement), yielding ICC = 0.87 (95% CI: 0.83–0.90) across all secondary indicators, indicating good to excellent agreement among annotators⁴⁸. Critically, data splitting was performed strictly at the teacher level: all observation cycles from a given teacher were assigned entirely to either the training set (80%) or the held-out test set (20%), ensuring that no information leakage occurred across splits. This teacher-level partitioning is essential because multiple observation cycles from the same teacher share idiosyncratic features that could inflate performance metrics if split across training and test sets.

The experimental design encompasses three evaluation threads: overall diagnostic performance comparison, modality contribution analysis through ablation, and dimension-specific diagnostic assessment. We benchmark our full multimodal VLM-based model (denoted VLM-Fusion) against five baseline and ablated variants. The evaluation metrics follow standard classification and regression conventions. For the discretized competency level classification task, continuous composite scores were mapped into five proficiency tiers using policy-based thresholds aligned with the national teacher competency standards issued by the Ministry of Education: Tier 1 (Novice, S ≤ 0.30), Tier 2 (Developing, 0.30 < S ≤ 0.50), Tier 3 (Competent, 0.50 < S ≤ 0.70), Tier 4 (Proficient, 0.70 < S ≤ 0.85), and Tier 5 (Expert, S > 0.85). The class distribution in our dataset is as follows: Tier 1 accounts for 8.3% (n = 32), Tier 2 for 24.1% (n = 93), Tier 3 for 34.7% (n = 134), Tier 4 for 22.8% (n = 88), and Tier 5 for 10.1% (n = 39), reflecting a roughly normal distribution with a slight positive skew characteristic of rural teacher populations. We report accuracy, macro-averaged precision, recall, and F1-score. For the continuous score prediction task, we adopt root mean squared error (RMSE) computed as:

20

where and are the predicted and ground-truth composite competency scores for the  n-th teacher. Additionally, to capture the rank-order fidelity of predictions—arguably more important than absolute accuracy for downstream path recommendation—we compute Spearman’s rank correlation coefficient between predicted and actual score rankings⁴⁹.

Table 5 presents the comprehensive performance comparison across all evaluated models, tested on a held-out set comprising 20% of the annotated teacher samples, stratified by province and school type to prevent geographic bias.

Table 5.

Diagnostic model performance comparison.

Model	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)	RMSE
Text-only (BERT)	71.3	69.8	68.5	69.1	0.142	0.761
Video-only (ViT)	66.7	64.2	65.9	65.0	0.168	0.714
Early fusion (Concat)	74.8	73.1	72.6	72.8	0.127	0.803
Late fusion (Avg.)	76.2	74.5	75.1	74.8	0.119	0.821
CLIP-based (no fine-tune)	78.4	77.0	76.3	76.6	0.108	0.847
VLM-fusion (w/o gating)	83.1	81.6	82.4	82.0	0.089	0.892
VLM-fusion (full)	86.5	85.2	84.7	84.9	0.074	0.918

All reported metrics represent the mean across five independent training runs with different random seeds. The full VLM-Fusion model achieves the highest performance across all metrics, reaching a mean F1-score of 84.9% ± 0.6% and a rank correlation of 0.918 ± 0.008 (95% confidence intervals: F1 [83.9%, 85.9%]; RMSE [0.069, 0.079]). The results in Table 5 reveal several noteworthy patterns—a substantial improvement over both unimodal baselines and simpler fusion strategies. The gap between text-only and video-only models is itself instructive: textual data from lesson plans and reflections proves more diagnostically informative than video alone, likely because written documents more directly encode the pedagogical reasoning that expert annotators weight heavily in their assessments. Yet the consistent gains from multimodal fusion confirm that visual evidence contributes complementary information that text cannot fully substitute.

Figure 3 illustrates the accuracy progression from unimodal through increasingly sophisticated fusion approaches, making the incremental benefit of each architectural choice visually apparent. The steepest gain occurs when moving from naive concatenation to cross-modal attention, underscoring the importance of interaction-aware fusion over simple feature stacking.

Fig. 3.

Overall diagnostic accuracy comparison across models.

To understand how diagnostic performance varies across competency dimensions, we disaggregated the F1-scores by primary dimension for both the full model and the strongest baseline (CLIP-based without fine-tuning).

As shown in Fig. 4, VLM-Fusion outperforms the baseline consistently, but the margin of improvement is notably larger for the Rural Contextual Adaptability dimension ()—an 11.3% point F1 gain—than for Pedagogical Content Knowledge (), where the gap narrows to 5.7 points. We attribute this to the fine-tuning process, which exposes the model to rural-specific visual and textual patterns largely absent from the generic pre-training corpus, thereby disproportionately benefiting dimensions that depend on context-sensitive features.

Fig. 4.

Dimension-specific diagnostic performance comparison between VLM-fusion and CLIP baseline.

The ablation study further clarifies the contribution of individual design components. We systematically removed the cross-modal attention module, the gating mechanism, and the supplementary modality streams (student feedback, metadata, self-assessment), measuring the F1-score drop in each condition.

Figure 5 presents the fusion strategy comparison, confirming that cross-modal attention yields the single largest performance increment among all fusion approaches tested. Removing it and reverting to late fusion reduces F1 by 10.1 points, a degradation that exceeds the impact of removing any single data modality entirely.

Fig. 5.

Impact of different modality fusion strategies on diagnostic performance.

Figure 6 depicts the ablation results more granularly. The gating mechanism contributes a 3.4-point F1 improvement—modest in absolute terms but disproportionately important under missing-modality conditions, where ungated models suffer an additional 8.2-point degradation compared to only 2.9 points for the gated variant⁵⁰. To quantify this robustness formally, we define a missing-modality resilience index:

21

Fig. 6.

Ablation experiment results comparing component contributions.

Our full model achieves an average  MRI=0.934 across single-modality dropout conditions, compared to 0.871 for the ungated variant, confirming that the adaptive gating design successfully insulates diagnostic quality against the data incompleteness scenarios that are routine in rural deployment contexts.

Diagnostic model performance evaluation

The aggregate results presented in Table 5 establish the overall superiority of VLM-Fusion, but a more penetrating analysis requires dissecting the contributions of individual architectural components and examining whether performance holds across demographically distinct teacher subgroups. We pursue both objectives in this subsection, beginning with a structured ablation study that isolates the marginal value of each module beyond the preliminary observations reported above.

Table 6 summarizes the ablation results across five model configurations, each obtained by removing a single component from the full VLM-Fusion architecture while keeping all other elements intact. Performance is reported in terms of accuracy, F1-score, RMSE, and the rank correlation defined previously.

Table 6.

Ablation experiment results.

Model configuration	Accuracy (%)	F1-Score (%)	RMSE
Full VLM-Fusion	86.5	84.9	0.074	0.918
w/o Cross-modal Attention	79.3	77.6	0.113	0.842
w/o Gating Mechanism	83.1	82.0	0.089	0.892
w/o Supplementary Modalities	84.2	82.8	0.085	0.896
w/o Domain Fine-tuning	80.7	79.4	0.105	0.858

As presented in Table 6, the cross-modal attention module contributes the largest single performance increment: its removal precipitates a 7.3-point drop in F1-score and inflates RMSE by over 50%. This confirms our earlier intuition that interaction-level fusion—where visual and textual features mutually inform one another before diagnostic inference—captures correlational structures that no amount of post-hoc combination can recover. Domain fine-tuning on rural educational data ranks as the second most impactful factor, yielding a 5.5-point F1 improvement over the frozen pre-trained backbone. This finding aligns with recent observations in the transfer learning literature that task-specific adaptation remains critical even for large-scale pre-trained models when the target domain departs substantially from pre-training distributions⁵¹.

To quantify the relative contribution of each component more precisely, we define a normalized contribution index that expresses each module’s impact as a proportion of the total performance gap between the weakest ablated variant and the full model:

22

where denotes the F1-score decrease when module  m is removed. Under this metric, cross-modal attention accounts for 39.7% of the total contribution, domain fine-tuning 29.9%, gating 15.8%, and supplementary modalities 14.6%—a distribution that should guide computational resource allocation in deployment scenarios where not all components can be maintained at full capacity.

The question of generalization deserves particular scrutiny, since rural teachers are far from a homogeneous population. Teaching experience, subject specialization, and geographic context all shape a teacher’s competency profile in ways that could interact unpredictably with model behavior. We partitioned the test set into four subgroups based on teaching experience: novice (0–3 years,  n=24), developing (4–8 years,  n=31), experienced (9–15 years,  n=16), and veteran (16 + years,  n=7).

Figure 7 depicts the F1-scores achieved by VLM-Fusion and the strongest baseline (CLIP without fine-tuning) across these four subgroups. Our model maintains F1-scores above 82% for all groups, though performance peaks for the developing cohort (87.1%) and dips modestly for veterans (82.4%). The baseline, by contrast, exhibits a sharper decline for both novice and veteran teachers—groups whose competency profiles deviate most from population averages. We hypothesize that the cross-modal attention mechanism helps disambiguate atypical cases by drawing on complementary evidence streams: a veteran teacher whose video data appears conventional may reveal distinctive reflective depth in written narratives, and vice versa. To test whether the performance differences across subgroups are statistically meaningful, we conducted a one-way ANOVA on per-teacher F1-scores across the four experience categories. Here, “per-teacher F1-score” is operationally defined as follows: for each teacher in the test set, we computed the F1-score by treating that teacher’s multiple observation-cycle predictions as a small classification set against the corresponding expert-rated tier labels, then averaged across cycles to obtain a single scalar per teacher. This yielded ,  p=0.083, which does not reach conventional significance thresholds. Before interpreting this result, we examined the parametric assumptions. Shapiro–Wilk tests for normality of per-teacher F1 within each subgroup returned p values of 0.19 (novice), 0.34 (developing), 0.11 (experienced), and 0.28 (veteran), suggesting no serious deviation from normality, although the veteran cohort is admittedly too small (n = 7) for the test to carry much power. Levene’s test for homogeneity of variance gave p = 0.37, which is consistent with equal variances across groups. As a robustness check against the assumption that per-teacher F1—averaged across only 2–4 observation cycles—is truly normally distributed at the individual level, we also ran a Kruskal–Wallis non-parametric test, which returned H(3) = 6.14, p = 0.105. The parametric and non-parametric verdicts therefore agree. We caution, however, that with subgroup sizes ranging from n = 7 (veterans) to n = 31 (developing), this analysis is substantially underpowered. Rather than report post-hoc observed power—a quantity that is, by now, widely recognized as a deterministic function of the p value and therefore informationally redundant—we frame the issue prospectively: a prospective power analysis conducted with G*Power 3.1 indicates that, to detect a medium effect (f = 0.25) with 80% power at α = 0.05 in a four-group ANOVA design, the total required sample size is approximately N = 180 (≈ 45 per group), which substantially exceeds our current subgroup sizes, and particularly the veteran cohort. The failure to reject the null should therefore be read as “not detected in this sample” rather than as “not present in the population,” and drawing firmer conclusions about cross-group equity will require targeted oversampling of underrepresented experience bands in future data collection⁵². Formally, we express the generalization stability criterion as:

23

Fig. 7.

Diagnostic performance comparison across teacher experience subgroups.

where is the F1-score for subgroup  g and is the overall mean. Our model achieves  GS=0.945, indicating that performance variation across teacher subgroups remains within 5.5% of the mean—a level of stability we consider encouraging, though we acknowledge that the GS metric defined in Eq. (23) is an ad hoc measure introduced for descriptive convenience in this study. It does not carry established psychometric properties and should be interpreted as a simple descriptive ratio rather than a validated fairness index. Future work could adopt more established group fairness metrics from the algorithmic fairness literature to substantiate equity claims more rigorously.

Learning path optimization effect verification

Diagnostic accuracy, however strong, means little if the paths generated from those diagnoses fail to produce tangible professional growth. To evaluate this downstream impact, we conducted a quasi-experimental field study over a 12-week period involving 96 rural teachers drawn from 23 schools that had not participated in the dataset construction phase described in Sect.  4.1. Participants were assigned to one of three conditions through stratified random assignment based on teaching experience and school resource level. We characterize this as a quasi-experimental design because, although assignment was randomized within strata, neither participants nor assessors were blinded to condition, and the sample was drawn from volunteer schools rather than a probability sample of the full rural teacher population. The three conditions were: an experimental group receiving fully adaptive path recommendations from our system ( n=34), an active control group following expert-curated but non-personalized training sequences ( n=32), and a self-directed group given access to the same learning resource library without any structured guidance ( n=30). All three groups had identical access to the underlying training materials—the sole difference lay in how resources were sequenced and recommended.

Competency was assessed at baseline (), midpoint (, week 6), and endpoint (, week 12) using the multimodal diagnostic procedure detailed in Sect.  3.1 and 3.2. We define the competency improvement ratio as the relative change in composite score over the study period:

24

The experimental group achieved a mean of 23.7% ( SD=6.4), compared to 14.2% ( SD=7.1) for the expert-curated control and 8.9% ( SD=8.3) for the self-directed group. A one-way ANCOVA controlling for baseline scores confirmed that these differences were statistically significant (,  p<0.001), with post-hoc Bonferroni comparisons revealing significant pairwise differences between the adaptive group and both control conditions ( p<0.01 in each case). The partial eta-squared effect size was 0.286, indicating a large effect. Adjusted group means (controlling for baseline) were 0.691 for the adaptive group, 0.588 for the expert-curated group, and 0.534 for the self-directed group. Homogeneity of regression slopes was verified (interaction p = 0.41), and Levene’s test confirmed equal variances (p = 0.23)⁵³. Regarding study integrity, 91 of the 96 enrolled teachers completed all three assessment points (attrition rate: 5.2%), with 2 dropouts from the adaptive group (both due to school transfer), 1 from the expert-curated group, and 2 from the self-directed group. Among completers, mean resource engagement rates were 87.3% in the adaptive group, 79.1% in the expert-curated group, and 61.4% in the self-directed group, measured as the proportion of recommended or available modules actually accessed. We detected no evidence of cross-group contamination, as participants were drawn from geographically separated schools. What surprised us, in retrospect, was not merely the magnitude of the improvement gap but its trajectory. The adaptive group’s gains accelerated between and , whereas the expert-curated group’s growth rate plateaued and the self-directed group’s actually decelerated—a pattern consistent with the hypothesis that ongoing diagnostic recalibration keeps recommendations aligned with evolving needs, preventing the stagnation that afflicts static curricula.

Figure 8 illustrates these divergent trajectories, plotting mean composite scores at each assessment point with 95% confidence intervals. The widening gap between the adaptive condition and both controls is visually striking, particularly in the second half of the study. Notably, variance within the adaptive group narrows over time—the system appears to be bringing initially diverse competency profiles toward more uniformly improved states, an equity-enhancing outcome that resonates with the broader motivations of this research.

Fig. 8.

Competency improvement trajectories across three experimental conditions over the 12-week study period.

Beyond raw competency gains, we assessed participant experience through a post-study survey administered to all 96 teachers, measuring perceived relevance of recommended activities, satisfaction with the learning experience, and self-reported time efficiency on five-point Likert scales. The adaptive group reported significantly higher satisfaction ( M=4.31,  SD=0.52) than the expert-curated ( M=3.67,  SD=0.71) and self-directed ( M=3.12,  SD=0.89) groups. Perceived time efficiency—a dimension we anticipated would be particularly salient for overburdened rural teachers—showed an even larger differential, with adaptive group participants rating the system 4.47 on average compared to 3.21 for the self-directed condition. To synthesize these multifaceted outcomes into a single effectiveness metric, we computed a weighted composite index:

25

where and denote normalized mean satisfaction and efficiency scores respectively, and , , reflect the primacy of actual competency improvement over subjective perception. The adaptive group’s composite  EI reached 0.812, compared to 0.594 for the expert-curated group and 0.428 for the self-directed condition. We offer this index as an exploratory synthesis rather than a validated effectiveness metric: the weighting scheme (0.5 / 0.25 / 0.25) is author-defined, its coefficients have not been externally calibrated against an established psychometric benchmark, and the three component measures are not independent in the strict statistical sense. Read in that spirit, the EI gradient is most usefully interpreted as a descriptive summary that complements—not substitutes for—the ANCOVA, satisfaction, and efficiency tests reported above, which together suggest (though do not definitively prove) that the proposed system delivers meaningful and well-received professional development support under authentic rural conditions⁵⁴.

Discussion

The experimental evidence assembled in Section IV paints a largely encouraging picture, but it would be premature to accept these results without probing their boundaries, examining what they reveal about underlying mechanisms, and confronting the practical complications that separate a research prototype from a deployed educational tool. We organize this discussion around four interlocking themes.

The first concerns what the VLM-based diagnostic approach genuinely adds—and where it falls short. The ablation results in Table 6 make clear that cross-modal attention is the single most consequential architectural choice, contributing nearly 40% of the total performance gain. This is not a trivial finding. It suggests that the diagnostic value of multimodal data lies less in the sheer volume of information and more in the interplay between modalities—the way a teacher’s written reflection can disambiguate a visually ambiguous classroom interaction, or the way a photograph of a cluttered blackboard can contextualize an otherwise competent-seeming lesson plan. A question worth examining more closely is why text appears more informative than video in this diagnostic context (Table 5: text-only F1 = 69.1% vs. video-only F1 = 65.0%). We suspect this reflects the nature of the expert annotation process itself: annotators weighted pedagogical reasoning and intentionality heavily, and these qualities are more directly encoded in written lesson plans and reflections than in visual behavioral patterns, which require substantial inference to connect to underlying competency. Video, however, contributes unique complementary signals—classroom management behaviors, student engagement cues, and physical resource utilization—that written documents cannot capture, which explains why multimodal fusion still substantially outperforms text alone. Regarding the disproportionate fine-tuning benefit for Rural Contextual Adaptability (Fig. 4 and 11.3-point gain versus 5.7 points for Pedagogical Content Knowledge), we attribute this to the distributional gap between pre-training data and rural educational content: generic VLMs have encountered abundant pedagogical content in their pre-training corpora but very little content depicting rural-specific practices such as multi-grade instruction, community-integrated lessons, or improvisation under resource constraints. Fine-tuning on our rural dataset therefore has the greatest marginal impact on precisely these underrepresented dimensions. At the same time, we must acknowledge that the system’s diagnostic accuracy, while strong in aggregate, remains imperfect in ways that matter. The 82.4% F1-score for veteran teachers, though not statistically distinguishable from other subgroups, hints at a ceiling effect: teachers with decades of experience often display competency patterns that are subtle, idiosyncratic, and poorly represented in the training data. Our model, trained predominantly on developing and mid-career teacher profiles, may lack sufficient exposure to the nuanced signatures of expert practice. This is a limitation we do not wish to minimize.

Second, the learning path optimization results deserve careful mechanistic interpretation. The accelerating improvement trajectory observed in the adaptive group (Fig. 8) contrasts sharply with the plateau in the expert-curated condition and, we believe, reflects a fundamentally different dynamic. Static curricula—even those designed by skilled experts—inevitably drift out of alignment with individual learner states as training progresses, because each teacher’s competency profile evolves along a unique trajectory that no predetermined sequence can anticipate. The closed-loop design of our system, where diagnostic reassessment triggers path recalibration at each step, maintains this alignment continuously. In essence, the system converts professional development from a one-time prescription into an ongoing conversation between the teacher’s evolving needs and the available resource ecology. The narrowing variance within the adaptive group further suggests an equity-enhancing property: the system appears to invest more intensively in teachers who start from lower baselines, gradually compressing the competency distribution rather than amplifying pre-existing gaps. The practical implication of this narrowing variance is significant for policy: it suggests that adaptive systems of this kind could serve as equity-enhancing tools that partially compensate for the uneven starting conditions among rural teachers, though we stress that a 12-week window is too brief to confirm whether this compression persists over longer time horizons or reverses once system support is withdrawn.

Third, we must confront the deployment challenges that any system of this kind faces in rural contexts. The most immediate is computational infrastructure. Our full model requires GPU-accelerated inference for real-time cross-modal attention computation—a resource unavailable at most rural schools. During the field study, we addressed this through a hybrid architecture where data was collected locally on mobile devices and transmitted to a cloud server for processing during off-peak hours. This asynchronous design worked adequately in our study context, but it introduces latency that could undermine user engagement if recommendation turnaround times stretch beyond what teachers perceive as responsive. Edge deployment of distilled model variants represents one promising mitigation path, though the performance trade-offs of model compression in this specific application domain remain to be quantified. A second, subtler challenge is teacher trust. Several participants in informal post-study interviews expressed uncertainty about why particular resources were recommended, highlighting a familiar interpretability gap. The system currently generates recommendations without accompanying explanations—an omission that future iterations should rectify through attention visualization or natural language rationale generation.

Finally, situating our work within the broader landscape of AI-assisted teacher development reveals both convergences and departures. Prior systems have demonstrated the feasibility of automated classroom observation analysis and text-based competency assessment as independent capabilities, but none—to our knowledge—has integrated these within a single closed-loop framework that feeds multimodal diagnosis directly into adaptive path planning. Our approach also differs from most existing learning path recommendation systems in its explicit modeling of resource accessibility constraints through the penalty terms in Eq. (17), a design choice motivated by the specific realities of rural education but potentially transferable to other resource-constrained professional development contexts. Where our work converges with prior findings is in confirming that personalization matters more than content quality alone: even when all groups accessed identical resources, the sequencing and timing of recommendations proved decisive, reaffirming a principle that the adaptive learning literature has long advocated but that teacher professional development research has been slow to operationalize at scale.

Conclusion

This paper has presented an integrated intelligent diagnostic system that fuses vision-language model capabilities with adaptive learning path optimization to address the persistent challenge of rural teacher professional development. The work spanned from theoretical grounding—constructing a four-dimensional competency model calibrated to rural teaching realities through AHP-weighted indicators—through system architecture design and algorithm development, to empirical validation via both offline benchmarking and a 12-week field experiment. What emerged is not merely a technical artifact but, we hope, a step toward rethinking how AI can serve educational equity in contexts that mainstream research too often overlooks.

The core contributions of this research can be distilled into three threads. First, the multimodal diagnostic framework demonstrated that cross-modal attention between visual and textual evidence streams yields substantially richer competency assessments than any single-modality or naive fusion approach. The 84.9% F1-score and 0.918 rank correlation achieved by the full VLM-Fusion model represent competitive results for automated teacher competency diagnosis in rural settings, though we note that direct comparison with prior work is complicated by the absence of a shared public benchmark in this domain. The gating mechanism proved particularly consequential for practical robustness—maintaining diagnostic quality even when individual modality streams were missing or degraded, a scenario that is the norm rather than the exception in under-resourced schools. Second, the Actor-Critic-based adaptive path generation algorithm, with its explicit modeling of time constraints and resource accessibility through the reward function in Eq. (17), produced measurably superior professional development outcomes compared to both expert-curated and self-directed alternatives. The 23.7% mean competency improvement in the adaptive group, against 14.2% and 8.9% in the control conditions, speaks to a fundamental insight: personalized sequencing matters as much as content quality. Third, the narrowing competency variance within the adaptive group provides a preliminary indication of a possible equity-enhancing dynamic—the system appears to invest proportionally more in teachers with greater initial deficits—though longer-term studies are needed to determine whether this compression persists and whether it translates into durable pedagogical improvement rather than short-term score gains.

The theoretical implications extend in two directions. For the educational technology community, this work demonstrates that contemporary VLMs, when appropriately fine-tuned and embedded within closed-loop systems, can transcend their origins in generic computer vision and natural language processing to address domain-specific diagnostic challenges that demand contextual sensitivity. For teacher education research more broadly, the results offer encouraging (though preliminary) empirical support for the long-standing argument that professional development should be treated as an adaptive, individualized process rather than a uniform curricular delivery.

We are candid about the limitations. The dataset, while carefully constructed, encompasses 386 teachers from five provinces—a sample that, however diverse within its scope, cannot claim to represent the full heterogeneity of rural education across an entire nation. The 12-week field study, though sufficient to detect meaningful competency gains, is too short to assess whether those gains persist or how the system performs over multiple academic years of continuous use. The model’s relative weakness on veteran teacher profiles, noted in Sect.  4.2, points to a data imbalance that targeted collection efforts could address but that we did not resolve in this iteration. Furthermore, the computational demands of real-time multimodal inference remain a barrier to fully autonomous deployment at the school level, currently necessitating cloud-based processing with attendant latency and connectivity dependencies.

These limitations naturally chart the course for future work. Expanding the dataset to encompass additional provinces and minority-language teaching contexts would strengthen generalizability claims. Longitudinal studies tracking teachers over two to three years would clarify whether adaptive path recommendations produce durable competency transformation or merely accelerate short-term gains. On the technical front, model distillation and edge deployment strategies deserve systematic investigation to reduce the infrastructure burden on rural schools. Perhaps most ambitiously, incorporating explainable AI mechanisms—generating human-readable justifications for both diagnostic assessments and path recommendations—could address the trust concerns surfaced in our field study and move the system from a black-box tool toward a genuine pedagogical partner. We close with a deliberately modest framing. What we present here is a promising, well-supported early-stage system demonstration, not a deployment-ready solution; the sample scope (386 teachers across five provinces), the short 12-week follow-up window, the underrepresentation of veteran teachers in the training data, and the current reliance on cloud-based inference all mark the distance that remains between this prototype and an infrastructure that could be recommended for broad adoption. If this work contributes to a future in which a teacher’s geographic circumstance no longer predetermines the quality of professional growth support available to them, it will have served its purpose.

Abbreviations

VLM

Vision-language model

AHP

Analytic hierarchy process

CLIP

Contrastive language-image pre-training

BLIP

Bootstrapped language-image pre-training

MDP

Markov decision process

RL

Reinforcement learning

GNN

Graph neural network

GAT

Graph attention network

PPO

Proximal policy optimization

DQN

Deep Q-network

ViT

Vision transformer

BERT

Bidirectional encoder representations from transformers

RMSE

Root mean squared error

MRI

Missing-modality resilience index

ANCOVA

Analysis of covariance

ANOVA

Analysis of variance

MLP

Multi-layer perceptron

CR

Consistency ratio

RI

Random consistency index

Author contributions

Lumei Wang conceptualized and designed the study, developed the theoretical framework and system architecture, implemented the VLM-based diagnostic model and adaptive learning path optimization algorithm, conducted the data collection and experimental evaluation, performed the statistical analysis, and drafted and revised the manuscript. The author read and approved the final manuscript.

Data availability

All data generated and analyzed during the current study are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

This study was approved by the Research Ethics Committee of Nanyang Institute of Technology (Reference Number: IRB-2024-03172). All participants provided written informed consent prior to enrollment. The study was conducted in accordance with the Declaration of Helsinki and relevant national regulations.

Consent for publication

All authors have reviewed the manuscript and consent to its publication. No identifiable information regarding participants has been included.