The research, conducted by Lin Cheng and colleagues at Universiti Putra Malaysia together with Riyan Hidayat of SEGi University, recruited 166 sixth-grade students aged 11 to 12 from two public primary schools in Guiyang, in China’s Guizhou Province. One group of 84 students learned six Tang poems through immersive VR instructional videos, while a control group of 82 students studied the same poems through traditional teacher-led lessons. Crucially, both groups followed identical learning objectives, schedules, and content, isolating the technology itself as the key difference between the classrooms.

The VR environment was not an off-the-shelf product but a carefully engineered instructional system built through the ADDIE model, the classic five-stage instructional design process of analysis, design, development, implementation, and evaluation. In the analysis phase, the team interviewed six experienced primary school Chinese teachers and reviewed 39 empirical studies, confirming that insufficient engagement, a lack of immersive scenarios, and a shortage of quality VR resources were the central obstacles to poetry teaching. The design stage produced a six-lesson program covering 48 core knowledge points, which experts rated with a content validity index of 0.97. The finished materials earned a mean expert rating of 4.85 out of 5.

Technically, the system ran on mobile phone-based head-mounted displays equipped with head-tracking and handheld controllers, allowing students to navigate panoramic poetic scenes from a first-person perspective. Rather than passively watching a video, learners freely explored virtual environments, inspected cultural elements from multiple angles, and engaged with embedded prompts and interactive recitation activities. Immersive visuals and 360-degree auditory experiences replaced the dry, abstract imagery that typically accompanies classical Chinese verses. Safety protocols were strict: continuous VR exposure was capped at 15 minutes followed by a five-minute rest, teachers monitored every session, and students could stop immediately if they experienced dizziness or visual discomfort. Headsets were sanitized before each use.

The study measured four outcomes: academic achievement, concentration, engagement, and classroom interaction. Achievement was assessed with a 12-item test combining multiple-choice, fill-in-the-blank, dictation, and short-answer questions scored against Bloom’s taxonomy, with internal consistency values of 0.81 and 0.82 for the pre- and post-tests. Concentration, engagement, and interaction were measured with validated Likert-scale questionnaires adapted from established instruments, all showing acceptable to strong reliability. To account for baseline differences between intact classes, the researchers used Analysis of Covariance, entering pre-test scores as a covariate, a statistically rigorous approach for a setting where random assignment of individual students was impractical.

The results were striking in places and nuanced in others. Students in the VR condition significantly outperformed their peers on the achievement post-test, with a small but reliable effect. Engagement showed the largest effect by far, with a substantial adjusted difference between groups, suggesting that the immersive experience moved students from passive memorization toward active emotional and behavioral involvement with the poems. Classroom interaction also improved significantly, though the effect was modest, likely because the intervention emphasized individual exploration over collaborative tasks. Concentration told a more complicated story: the VR group scored higher descriptively, but the standard statistical test found no significant main effect, and an assumption violation in the model means those results must be interpreted with caution.

The authors ground these findings in established theory. The Cognitive-Affective Model of Immersive Learning holds that presence, agency, embodiment, and cognitive engagement drive learning outcomes in VR, while Flow theory describes the state of complete absorption that immersive environments can induce. Motivation and engagement theory adds that heightened intrinsic motivation feeds directly into behavioral and emotional engagement, which in turn supports achievement. The study is careful to note that these frameworks served as conceptual scaffolding rather than directly measured constructs, and it acknowledges that some VR benefits in prior research may stem from novelty effects that fade with repeated exposure.

What makes this study notable is its target domain. Educational VR research has been overwhelmingly concentrated in STEM subjects, where virtual laboratories and geometric visualizations are natural fits. Tang poetry presents a harder problem: condensed classical language, dense historical symbolism, and imagery that demands imaginative reconstruction. By demonstrating measurable gains in achievement and engagement in a humanities context, the study extends the evidence base for immersive technology well beyond science classrooms and offers curriculum developers a concrete framework, structured around ADDIE, for integrating VR into language and literature instruction.

The researchers are candid about the limitations. The sample came from just two schools in one Chinese city, the six-week intervention was too brief to reveal long-term effects, and the interaction questionnaire captured students’ perceptions of teacher support rather than peer-to-peer exchange. The head-mounted displays also housed smartphones from different manufacturers, including Huawei and Xiaomi models, so hardware variation may have influenced perceived immersion. Formal inter-rater reliability for the short-answer scoring was not assessed, a limitation the authors themselves flag. They call for randomized and multi-site designs, longer intervention cycles, qualitative methods such as interviews and classroom observations, and comparative studies of semi-immersive versus fully immersive setups.

Even with those caveats, the implications ripple outward. As Chinese language learning surges globally and schools everywhere experiment with immersive technology, this study provides empirical evidence that VR can carry cultural and literary weight, not just scientific visualization. The authors recommend improved VR infrastructure, stronger teacher training, and closer alignment of VR resources with curriculum standards, along with collaborative tasks and gamified elements to strengthen the social dimension of future implementations. For now, the image of eleven-year-olds in Guiyang stepping inside a Tang dynasty landscape, reciting verses as they explore them, offers a vivid glimpse of how classical literature might survive its collision with the digital age, and perhaps be revitalized by it.

Subject of Research: The effect of immersive virtual reality instruction on primary school students' learning of Tang poetry

Article Title: Immersive virtual reality supports Tang poetry learning among primary school students

Article References: Immersive virtual reality supports Tang poetry learning among primary school students. (n.d.). https://doi.org/10.1007/s44217-026-02229-8

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02229-8

Keywords: virtual reality, Tang poetry, primary education, Chinese language learning, engagement, classroom interaction, ADDIE model, quasi-experimental design, educational technology, immersive learning, humanities education, ANCOVA

In a paper published in the International Journal of Data Science and Analytics, Yang Yang of Zhengzhou Railway Vocational and Technical College and Henan Agriculture University, and Di Zhao of Henan Open University and Zhengzhou University, describe an adaptive teaching algorithm that combines natural language processing (NLP) with deep reinforcement learning (DRL) to tutor students through Chinese grammar rules. The system does not simply detect errors and dispense corrections. Instead, it continuously chooses teaching strategies based on an evolving profile of each learner, adjusting the difficulty of exercises in real time and tailoring feedback to the individual’s demonstrated weaknesses and progress.

The motivation behind the work stems from a persistent pedagogical frustration. Rule-based teaching approaches, the authors note, tend to be inflexible: they treat grammar acquisition as a fixed sequence of rules to be drilled, regardless of whether a particular learner has already mastered a rule or is struggling with its prerequisites. That one-size-fits-all structure can make learning inefficient and demotivating, particularly for Chinese, where multiple linguistic properties intersect — a mismatch the researchers identify as the central obstacle the study set out to address.

Technically, the system rests on two cooperating components. The NLP layer is responsible for grammar itself: parsing learner sentences, recognizing grammatical errors and encoding rule structures so that the system understands what a student has written and where it deviates from standard usage. The DRL layer then operates on top of that analysis as a decision-making engine. Framed as a reinforcement learning problem, the tutoring process becomes a sequence of states (the learner’s current profile and performance history), actions (available teaching strategies, feedback types and difficulty levels) and rewards (signals tied to learning gains and error correction outcomes). The agent — in this case, the tutoring algorithm — learns a policy for selecting the teaching action most likely to improve a given student’s grammar competence at that moment.

What distinguishes the proposed model from more conventional AI tutoring architectures is its specific choice of learning machinery. Rather than relying on a fixed algorithm whose behavior is static once trained, the researchers incorporate proximal policy optimization, or PPO, a policy-gradient reinforcement learning method known for balancing exploration of new teaching strategies against stable exploitation of strategies already known to work. PPO’s clipped update mechanism constrains how drastically the policy can change from one training iteration to the next, which helps prevent the catastrophic policy collapse that can plague more aggressive reinforcement learning methods. The system is further stabilized by experience replay, a technique in which past interactions between the learner and the system are stored and resampled during training, allowing the algorithm to learn from a broader distribution of past episodes rather than only the most recent ones. Together, the authors argue, these components give the model both the adaptability to respond to individual learners and the stability to keep improving without erratic swings in behavior.

The experimental results reported in the paper are striking. When compared with traditional rule-based teaching methods, the DRL-based adaptive system reduced the average time to mastery of grammar content from 12 sessions to 4 — a 66.7 percent improvement. Grammar error correction accuracy rose from 78.2 percent to 95.3 percent, a gain of 21.8 percentage points. Beyond those headline numbers, the system produced more consistent learning progression across sessions, an effect the researchers attribute to the combination of personalized feedback and adaptive difficulty adjustment. Students were not simply moving faster; their trajectories were smoother, with fewer of the stalls and regressions that characterize one-size-fits-all grammar instruction.

The comparison to other AI-based grammar correction systems is also favorable, according to the study. Relative to existing AI-ready grammar correction tools, the proposed method demonstrated greater flexibility, greater effectiveness, and higher accuracy in correcting grammatical errors. The authors position this as a meaningful advance over prior work in Chinese grammatical error correction, a field that has seen considerable activity — including syntax-enhanced correction models, fine-grained error corpora and alignment-enhanced correctors — but which has largely focused on detecting and fixing errors rather than on deciding, pedagogically, what a learner should do next. The new system’s contribution lies in that decision layer: the reinforcement learning policy that treats teaching itself as a learnable optimization problem.

The implications extend beyond Chinese language classrooms. The architecture — NLP for domain analysis, DRL for pedagogical policy — is in principle language-agnostic, and the authors describe the work as providing an original, data-informed approach to intelligent grammar learning that can scale to adaptive language learning in AI-supported environments more broadly. Interest in AI-driven language tutoring has accelerated sharply in recent years, with studies examining everything from ChatGPT’s effect on foreign language learners to AI systems that facilitate cross-cultural communication and learner-generated contexts. Much of that work, however, has centered on English, whose massive corpora and mature NLP tooling make it the default testbed for educational AI. Chinese has been comparatively underserved, in part because its grammar resists the kind of explicit rule enumeration that powers many English tutoring systems. A framework that learns teaching strategy dynamically rather than encoding it statically may be better suited to such languages.

The study arrives amid a broader shift in how educational technologists think about personalization. Earlier generations of adaptive learning platforms relied largely on pre-defined learner models and hand-crafted rules for content sequencing. Reinforcement learning flips that paradigm: instead of a designer deciding in advance what the best next exercise is for a student archetype, the system discovers the policy empirically, guided by reward signals that reflect actual learning outcomes. The inclusion of PPO and experience replay in this study suggests a maturing understanding within the educational AI community that the training dynamics of the algorithm matter as much as the final policy — an unstable tutor that improves for one learner while regressing for another is worse than no tutor at all.

There are, of course, caveats worth keeping in view. The paper reports that no datasets were generated or analyzed during the current study as deposited artifacts, and the details of learner populations, evaluation protocols and long-term retention effects would need to be examined in the full text for educators weighing deployment. Reinforcement learning systems also demand careful reward design: if the reward signal rewards short-term error reduction over durable understanding, an adaptive tutor could optimize toward superficial pattern-matching rather than genuine grammatical competence. The authors’ emphasis on consistent progression across sessions suggests they have attended to this concern, but it remains one of the defining challenges of the field.

The work was funded in part by the Henan Province Science and Technology Research Project on key technologies for optimizing the spatial layout of complex terrain farmland based on deep learning — an indication of how deep learning expertise cultivated in one application domain can be redirected toward educational technology. Yang led conceptualization, methodology, software development, data curation, formal analysis and original drafting, while Zhao supervised the project, handled validation and administered the work, according to the paper’s contribution statement.

For the millions of students worldwide studying Chinese as a second language — a population that has grown alongside China’s economic and cultural influence — the prospect of a tutor that adapts its strategy session by session, identifies exactly which grammar points need reinforcement, and calibrates difficulty to keep learners in the productive zone between boredom and frustration, represents a tangible near-future improvement over static workbooks and uniform lesson plans. For researchers, the study adds a data point to a growing argument that reinforcement learning, long confined to games and robotics, has a genuine role to play in the delicate, human-centered problem of teaching. Whether such systems can match the judgment of skilled human instructors in edge cases remains an open question, but the measured gains reported here — cutting time to mastery by two-thirds while pushing correction accuracy above 95 percent — make a compelling case that the answer, for at least some dimensions of grammar instruction, is moving steadily toward yes.