Pith. sign in

REVIEW 4 major objections 5 minor 53 references

A curriculum-aligned AI system for Vietnamese classrooms reports 87% accuracy on the national high school exam and slashes lesson planning from 3–4 hours to roughly 45 minutes.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A Vietnamese education LLM (Qwen3-8B + SFT/DPO) with an ADDIE-style lesson-planning workflow reports 87% exam accuracy and 30–45 min prep, with small satisfaction surveys.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection Solid engineering paper on a Vietnamese AI lesson-planning system; the headline accuracy figure is likely inflated by possible data leakage the authors acknowledge, and the prep-time numbers are inconsistent, but the system design is thoughtful and the paper is honest about its limits. the 4 major comments →

arxiv 2607.28647 v1 pith:ZBC6QLFK submitted 2026-06-02 cs.HC cs.AI

ConnectED: A Curriculum-Aligned AI System for Vietnamese Instructional Lesson Planning and Student Learning

Thang Doan Viet , Anh Nguyen Hoang , Tinh Luong Son , Anh Hoang Thi Ngoc , Huyen Giang Thi Thu , Tai Le Quy This is my paper
classification cs.HC cs.AI
keywords AI in EducationHuman-centered AIEducational LLMDPOADDIECurriculum-aligned lesson planningVietnamese educationTeacher-in-the-loop
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper presents ConnectED, a human-centered AI system that supports the full Vietnamese instructional lifecycle: curriculum-compliant lesson planning, interactive student practice, and feedback-driven revision. Its central claim is that a Vietnamese educational language model (VietEduQwen, fine-tuned from Qwen3-8B with supervised fine-tuning and direct preference optimization) reaches 87.02% accuracy on the 2025 national high school examination, 6.10 points above the base model. The authors further claim that structuring generation through the ADDIE instructional-design framework, with a teacher validation gate at every phase boundary, improves compliance and teacher trust compared with unstructured prompting, and that this workflow cuts lesson preparation from 3–4 hours to roughly 30–45 minutes. A sympathetic reader would care because the system addresses a concrete regulatory and pedagogical context rather than a generic benchmark, showing that orchestration and alignment, not just model scale, can drive practical gains.

Core claim

ConnectED operationalizes the ADDIE instructional design model as an orchestration layer for LLM-based lesson generation. Each of ADDIE's five phases — Analysis, Design, Development, Implementation, Evaluation — is encoded as a structured prompt that embeds Vietnam's Official Dispatch 5512 requirements and doubles as a teacher validation checkpoint. VietEduQwen, an 8-billion-parameter Vietnamese educational model trained with SFT and DPO, serves as the generation engine. The central discovery is that this combination yields a system more accurate and more trusted than the same model used with unstructured prompting: the DPO-trained model answers 87.02% of national exam questions correctly, a

What carries the argument

ADDIE — a five-phase instructional design model (Analysis, Design, Development, Implementation, Evaluation) — used as an orchestration layer. The paper's key move is to treat each phase as both a generation step and a teacher validation gate, encoding Official Dispatch 5512 requirements into structured prompt templates. This turns a generic LLM into a curriculum-compliant lesson generator, and the Evaluation phase connects student learning signals back to the Design and Develop phases. A secondary mechanism is VietEduQwen, a Vietnamese educational LLM fine-tuned from Qwen3-8B via SFT and DPO, which supplies the underlying reasoning and generation; ablations show both the DPO training and the

Load-bearing premise

The load-bearing premise is that the 3,119-question 2025 national high school exam used for evaluation was not present in VietEduQwen's training corpus; the authors themselves state that the possibility of data leakage between training and evaluation benchmarks cannot be entirely excluded.

What would settle it

Run VietEduQwen on the next national high school exam (e.g., 2026, released after the training-data cutoff) and compare accuracy; a drop well below 87% would indicate the 2025 benchmark was contaminated. A simpler check: compute n-gram overlap or run membership-inference tests between the training corpus and the 3,119 evaluation questions.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If VietEduQwen's 87.02% accuracy holds, a sub-10B open-weight model can match or approach frontier commercial models on Vietnamese national exam subjects, making high-quality educational AI feasible in school settings with limited compute.
  • The ADDIE orchestration pattern is claimed to contribute independent value: replacing it with unstructured prompting degrades both preparation efficiency and teacher satisfaction (p<0.01).
  • Removing the teacher-in-the-loop stage or the curriculum checker each causes a statistically significant drop in teacher satisfaction, indicating teachers value oversight even at the cost of extra time.
  • The student-facing Playground collects structured learning signals that feed back into lesson refinement, forming a closed loop without requiring new assessment infrastructure.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Inference: The ADDIE-as-orchestration pattern likely transfers to other regulated document-generation settings — e.g., individualized education plans, clinical teaching plans, or compliance-heavy training materials — wherever phase boundaries map naturally to approval checkpoints.
  • Inference: The 6.10-point gain attributed to DPO+SFT suggests domain-specific preference tuning on a modest base model is an effective alternative to scaling; a natural test is to apply the same recipe to another national curriculum and check whether the gain replicates.
  • Inference: The teacher-in-the-loop results imply adoption hinges on trust and workflow fit at least as much as raw accuracy, since satisfaction drops significantly when gates are removed.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper presents ConnectED, a teacher-in-the-loop AI system for Vietnamese secondary education, built on VietEduQwen, an SFT+DPO fine-tuned Qwen3-8B model. The system operationalizes the ADDIE instructional-design framework with structured prompts aligned to Official Dispatch 5512, adds a multi-agent Manim-based STEM visualization pipeline, and includes a student-facing Playground that feeds learning signals back to lesson refinement. The central quantitative claims are (i) 87.02% accuracy on 3,119 questions from the 2025 Vietnamese National High School Examination, a +6.10 percentage-point gain over Qwen3-8B; (ii) reduction of lesson preparation time from 3–4 hours to 30–45 minutes (or 15–20 minutes in Table 2); and (iii) high teacher/student satisfaction in small surveys. Ablations are used to attribute gains to DPO and to ADDIE orchestration.

Significance. If the quantitative claims were fully substantiated, the paper would be a useful contribution to AI-assisted education in a specific regulatory context: it combines a domain-tuned model, curriculum-aware orchestration, teacher validation checkpoints, and a student feedback loop, and its system-level ablation is a thoughtful way to isolate the orchestration layer. The authors also deserve credit for explicitly limiting claims to perceived utility rather than causal learning gains, and for making prompt templates available. However, the headline accuracy rest on a benchmark that overlaps with the training corpus, and the reported efficiency figures are inconsistent. These issues are load-bearing because the abstract and conclusions advertise the accuracy gain and the preparation-time reduction as the main empirical results.

major comments (4)
  1. [§3.1 and §5] The training corpus includes 'national examination materials, including official textbooks for students and teachers' (§3.1), and §5 concedes that 'the possibility of data leakage between training and evaluation benchmarks cannot be entirely excluded.' Since the evaluation is on the 2025 National High School Examination, this is not a peripheral caveat: any direct or near-duplicate 2025 exam items in training would inflate both the 87.02% figure and the +6.10pp gain over Qwen3-8B. The paper provides no contamination audit, no temporal separation, and no release of training data or model weights. Please add an exact/near-duplicate search against the training corpus, report accuracy on a strictly temporally disjoint set, or otherwise demonstrate that the evaluation is uncontaminated. Without this, the headline claim is not interpretable.
  2. [Table 2 vs Table 4 and Abstract] The reported preparation time is internally inconsistent. The Abstract and Table 4 report 30–45 minutes for the full ConnectED pipeline, while Table 2 reports 15–20 minutes and a 9–12× speedup; the Conclusion instead claims a 5–8× reduction (from 3–4 hours, 30–45 minutes gives 4–8×). This is the second headline quantitative claim, and it is not measured consistently. Please specify exactly what is being timed (including teacher review and revision), report the protocol under which each number was collected, and reconcile the discrepancy across abstract, tables, and conclusion.
  3. [§4.2, Table 1; §4.4, Table 3] All accuracy comparisons are reported as point estimates with no confidence intervals or significance tests. For n=3,119 questions, the standard error of an 87% accuracy is roughly 0.6 percentage points, so a 1.11pp DPO-ablation difference (Table 3) may or may not be meaningful. The Wilcoxon p-values in Table 4 are reported for satisfaction, but the accuracy table lacks equivalent treatment. Please report CIs, a significance test for the DPO ablation, and per-subject uncertainty in Fig. 7. This is necessary to support the quantitative claims, especially the +6.10pp gain.
  4. [§4.3 and §4.4] The user-satisfaction evidence rests on 18 teachers and 214 students with no demographic or sampling details, and the teacher sample in particular is small. While the paper appropriately avoids causal learning claims, statements such as 'strong acceptance' and '94% willingness to reuse' should be accompanied by response-rate information, inclusion criteria, and a discussion of how representative the sample is. This is secondary to the accuracy and prep-time concerns, but it affects the support for the system-level claims.
minor comments (5)
  1. [§3.2] Typo: 'structured promptsather than hardcoded rules' should be 'structured prompts rather than hardcoded rules.'
  2. [Fig. 7] The figure includes 'Civic-education' but the text in §4.2 lists eight subjects as Mathematics, Physics, Chemistry, Biology, English, Literature, Geography, and History. Please align the figures and text, and clarify the complete subject list.
  3. [§3.1, Eq. (2)] The SFT regularization weight λ in Eq. (2) is not defined or reported. Similarly, β in Eq. (1), the number of epochs, and the Wikipedia filtering threshold are not given numerically. Reporting these hyperparameters would improve reproducibility.
  4. [Throughout] The paper says '6% improvement' in §3.1 and '6.10% accuracy gain' in the Conclusion, while Table 1 correctly reports a 6.10 percentage-point gain. Please use percentage points consistently.
  5. [§4.1] The error analysis says 'a random sample of 200 errors' but does not report inter-annotator agreement for the manual categorization. Please provide agreement statistics or clarify the annotation procedure.

Circularity Check

0 steps flagged

No circularity found: the accuracy claim is an external benchmark measurement, and the acknowledged training/evaluation overlap is a data-contamination risk, not a within-paper circular derivation.

full rationale

The paper's central quantitative claim is the 87.02% accuracy of VietEduQwen on 3,119 questions from the 2025 Vietnamese National High School Examination. This is an empirical measurement against an external benchmark, not a quantity derived from fitted parameters or from the paper's own prior results. The DPO preference signal comes from Gemini 3 Pro as an independent judge, and the SFT/DPO loss functions are standard and external to the evaluation. The ADDIE orchestration claim is supported by a system-level ablation comparing ConnectED against unstructured prompting of the same model, which is an empirical comparison rather than a tautology. The acknowledged limitation in Section 5—'the possibility of data leakage between training and evaluation benchmarks cannot be entirely excluded'—is a legitimate validity threat to the 87.02% figure, and the reported preparation-time inconsistency between Table 2 (15–20 min) and Table 4 (30–45 min) is a reliability concern. However, neither reduces the paper's derivation to its own inputs by definition or by self-citation. No circular step can be quoted and exhibited. The appropriate finding is therefore no significant circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central accuracy claim depends on the training/evaluation data split and on unstated hyperparameters; the workflow claims depend on self-reported satisfaction and prep-time logs. No fundamentally new entities are introduced.

free parameters (4)
  • β (DPO preference-margin temperature)
    Controls sharpness of preference margin in Eq. (1); value not reported.
  • λ (SFT regularization weight)
    Balances DPO and SFT losses in Eq. (2); value not reported.
  • Number of fine-tuning epochs = 4
    Reported accuracy increases monotonically through Epoch 4; no held-out validation set is described, so the checkpoint appears to be selected on the evaluation benchmark, which can overstate accuracy.
  • Wikipedia article filtering threshold = 700 of 2,600
    Curation choice for training data; effect on results not analyzed.
axioms (5)
  • standard math DPO loss (Eq. 1) and combined DPO+SFT loss (Eq. 2) are valid and correctly implemented with the stated preference triples.
    The paper relies on Rafailov et al.'s DPO formulation without re-deriving it.
  • ad hoc to paper The 3,119-question 2025 exam set is a fair, uncontaminated benchmark.
    Its construction is not documented and the authors concede leakage cannot be excluded (Section 5).
  • domain assumption Official Dispatch 5512 requirements are correctly encoded in the prompt templates and teacher satisfaction reflects curriculum compliance.
    No independent content audit is provided; n=18 teacher survey is the only evidence.
  • domain assumption Survey respondents (18 teachers, 214 students) are representative and self-reports are accurate.
    Voluntary convenience sample; no incentives or response-rate data.
  • domain assumption In-system logs accurately capture end-to-end lesson preparation time.
    Prep-time numbers vary across the paper (15–20 vs 30–45 min), so log definitions are unclear.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of ConnectED: A Curriculum-Aligned AI System for Vietnamese Instructional Lesson Planning and Student Learning." pith.science (2026). https://pith.science/paper/ZBC6QLFK

@misc{pith2026260728647,
  author       = {Pith},
  title        = {Pith review of: ConnectED: A Curriculum-Aligned AI System for Vietnamese Instructional Lesson Planning and Student Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZBC6QLFK}},
  note         = {Machine review of arXiv:2607.28647}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This paper presents ConnectED, a human-centered AI system that supports the full instructional lifecycle in Vietnamese education by linking curriculum-aligned lesson design, interactive student learning, and feedback-driven refinement. Built on VietEduQwen, a Vietnamese educational large language model trained via supervised fine-tuning and direct preference optimization, the system ensures academically accurate, pedagogically appropriate, and student-safe interactions. ConnectED operationalizes the ADDIE framework through structured prompt templates aligned with Official Dispatch No. 5512/BGDDT-GDTrH, where each phase serves as both a generation step and a teacher validation gate. The Evaluation phase further closes the loop by connecting student performance data with iterative lesson improvement. Beyond lesson generation, the system integrates a student-facing interactive environment, enabling continuous collection of learning signals to support teacher decision-making. Evaluation on 3,119 questions from the 2025 Vietnamese National High School Examination shows that VietEduQwen achieves 87.02% accuracy, outperforming Qwen3-8B by 6.10 percentage points. Surveys of teachers (n=18) and students (n=214) demonstrate strong satisfaction with curriculum alignment, lesson clarity, and usability. In practice, lesson preparation time is reduced from 3--4 hours to approximately 30--45 minutes with teacher-in-the-loop review. Ablation studies confirm that both DPO training and ADDIE-based orchestration contribute independently to system performance, highlighting the importance of structured teacher oversight for practical deployment.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

53 extracted references · 10 canonical work pages · 2 internal anchors

  1. [1]

    5512/BGDÐT-GDTrH dated December 18, 2020 on the development and im- plementation of the school education plan

    The Ministry of Education and Training of Vietnam: Official Dispatch No. 5512/BGDÐT-GDTrH dated December 18, 2020 on the development and im- plementation of the school education plan. Accessed January 21, 2026 (2020). https://thuvienphapluat.vn/cong-van/Giao-duc/Cong-van-5512-BGDDT-G DTrH-2020-xay-dung-va-to-chuc-thuc-hien-ke-hoach-giao-duc-cua-nha-truon ...

    2020

  2. [2]

    134/2025/QH15 of the National Assem- bly: Law on Artificial Intelligence

    National Assembly of Vietnam: Law No. 134/2025/QH15 of the National Assem- bly: Law on Artificial Intelligence. Accessed January 21, 2026 (2025). https://va nban.chinhphu.vn/?pageid=27160&docid=216334&classid=1&typegroupid=3

    2025

  3. [3]

    International Journal of Modern Education and Computer Science 15(6), 14–30 (2023) https://doi.org/10.5815/ijmecs.2023.06.02

    Dao, X.-Q., Le, N.-B.: Llms performance on vietnamese high school biology ex- amination. International Journal of Modern Education and Computer Science 15(6), 14–30 (2023) https://doi.org/10.5815/ijmecs.2023.06.02

    work page doi:10.5815/ijmecs.2023.06.02 2023

  4. [4]

    VNHSGE: VietNamese High School Graduation Examination Dataset for Large Language Models

    Dao, X.-Q., Le, N.-B., Vo, T.-D., Phan, X.-D., Ngo, B.-B., Nguyen, V.-T., Nguyen, T.-M.-T., Nguyen, H.-P.: Vnhsge: Vietnamese high school graduation examination dataset for large language models. arXiv:2305.12199 (2023) https: //doi.org/10.48550/arXiv.2305.12199

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2305.12199 2023

  5. [5]

    In: Proceedings of the 38th Pacific Asia Conference on Language, Information and Computation, pp

    Do Minh, D., Van, V.N., Cong, T.D.: Using large language models for educa- tion managements in vietnamese with low resources. In: Proceedings of the 38th Pacific Asia Conference on Language, Information and Computation, pp. 20–34 (2024).https://aclanthology.org/2024.paclic-1.3.pdf

    2024

  6. [6]

    In: International Conference on Cooperative Design, Visualization and Engineering, pp

    Anh, N.T., Hung, P.D.: Development of vietnamese large language model for elementary math solving problem. In: International Conference on Cooperative Design, Visualization and Engineering, pp. 73–82 (2024). https://doi.org/10.100 7/978-3-031-71315-6_8 . Springer

    2024

  7. [7]

    In: JSAI International Symposium on Artificial Intelligence, pp

    Le, N.-K., Nguyen, D.-H., Do, D.-T., Nguyen, C., Nguyen, M.L.: Vietnamese ele- mentary math reasoning using large language model with refined translation and dense-retrieved chain-of-thought. In: JSAI International Symposium on Artificial Intelligence, pp. 260–268 (2024). https://doi.org/10.1007/978-981-97-3076-6_18 . Springer

    work page doi:10.1007/978-981-97-3076-6_18 2024

  8. [8]

    1–6 (2025)

    Thanh, T.L.T., Trong, H.H., Gia, B.T., Nguyen, T., Do, T., Ngo, T.D., Nguyen, N.L.-T.: Context-aware question answering for vietnamese university admissions viamulti-llmarchitecture.In:2025InternationalConferenceonMultimediaAnal- ysis and Pattern Recognition (MAPR), pp. 1–6 (2025). https://doi.org/10.1109/ MAPR67746.2025.11133885 . IEEE

    arXiv 2025

  9. [9]

    GreenMind: A Next-Generation Vietnamese Large Language Model for Structured and Logical Reasoning

    Tung, L.Q., Viet, H.Q., Loc, P.B., Thu, V.T.: Greenmind: A next-generation vietnamese large language model for structured and logical reasoning. arXiv:2504.16832 (2025) https://doi.org/10.48550/arXiv.2504.16832

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2504.16832 2025

  10. [10]

    Humanities and So- cial Sciences Communications12(1), 1784 (2025) https://doi.org/10.1057/s415 99-025-06004-2

    Zheng,Y.,Huang,S.,Zeng,X.,Huang,Y.,Liu,Z.,Luo,W.:Knowledge-enhanced large language models for automatic lesson plan generation. Humanities and So- cial Sciences Communications12(1), 1784 (2025) https://doi.org/10.1057/s415 99-025-06004-2

    work page doi:10.1057/s415 2025

  11. [11]

    Branch, R.M., Varank, İ.: Instructional Design: The ADDIE Approach vol. 722. Springer, New York (2009). https://doi.org/10.1007/978-0-387-09506-6 19

    work page doi:10.1007/978-0-387-09506-6 2009

  12. [12]

    Learning and individual differences103, 102274 (2023) https://doi.org/10.101 6/j.lindif.2023.102274

    Kasneci, E., Seßler, K., Küchemann, S., Bannert, M., Dementieva, D., Fischer, F., Gasser, U., Groh, G., Günnemann, S., Hüllermeier, E.,et al.: Chatgpt for good? on opportunities and challenges of large language models for education. Learning and individual differences103, 102274 (2023) https://doi.org/10.101 6/j.lindif.2023.102274

    arXiv 2023

  13. [13]

    British Journal of Educational Technology 55(1), 90–112 (2024) https://doi.org/10.1111/bjet.13370

    Yan, L., Sha, L., Zhao, L., Li, Y., Martinez-Maldonado, R., Chen, G., Li, X., Jin, Y., Gašević, D.: Practical and ethical challenges of large language models in ed- ucation: A systematic scoping review. British Journal of Educational Technology 55(1), 90–112 (2024) https://doi.org/10.1111/bjet.13370

    doi:10.1111/bjet.13370 2024

  14. [14]

    Educational Technology & Society28(2), 327–347 (2025)

    Tao, L., Deng, H., Song, Y.: Generative artificial intelligence in education. Educational Technology & Society28(2), 327–347 (2025)

    2025

  15. [15]

    In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pa- pers), pp

    Vanzo, A., Chowdhury, S.P., Sachan, M.: Gpt-4 as a homework tutor can improve student engagement and learning outcomes. In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pa- pers), pp. 31119–31136 (2025). https://doi.org/10.18653/v1/2025.acl-long.15 02

    work page doi:10.18653/v1/2025.acl-long.15 2025

  16. [16]

    In: International Conference on Interactive Collaborative Learning, pp

    Modran, H.A., Bogdan, I.C., Ursutiu, D., Samoilă, C., Modran, P.L.: Llm in- telligent agent tutoring in higher education courses using a rag approach. In: International Conference on Interactive Collaborative Learning, pp. 589–599 (2024). https://doi.org/10.1007/978-3-031-83520-9_54 . Springer

    work page doi:10.1007/978-3-031-83520-9_54 2024

  17. [17]

    In: Findings of the Association for Computational Linguistics: EMNLP 2020, pp

    Nguyen, D.Q., Nguyen, A.-T.: Phobert: Pre-trained language models for viet- namese. In: Findings of the Association for Computational Linguistics: EMNLP 2020, pp. 1037–1042 (2020). https://doi.org/10.18653/v1/2020.findings-emnlp.9 2

    doi:10.18653/v1/2020.findings-emnlp.9 2020

  18. [18]

    In: Proceedings of the 2022 Con- ference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Student Research Workshop, pp

    Phan, L., Tran, H., Nguyen, H., Trinh, T.H.: Vit5: Pretrained text-to-text transformer for vietnamese language generation. In: Proceedings of the 2022 Con- ference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Student Research Workshop, pp. 136–142 (2022). https://doi.org/10.18653/v1/2022.naacl-srw.18

    work page doi:10.18653/v1/2022.naacl-srw.18 2022

  19. [19]

    arXiv preprint arXiv:2311.02945 (2023) https://doi.org/10.48550/arXiv.2311.02945

    Nguyen,D.Q.,Nguyen,L.T.,Tran,C.,Nguyen,D.N.,Phung,D.,Bui,H.:Phogpt: Generative pre-training for vietnamese. arXiv preprint arXiv:2311.02945 (2023) https://doi.org/10.48550/arXiv.2311.02945

    Pith/arXiv arXiv doi:10.48550/arxiv.2311.02945 2023

  20. [20]

    In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pp

    Nguyen, X.-P., Zhang, W., Li, X., Aljunied, M., Hu, Z., Shen, C., Chia, Y.K., Li, X., Wang, J., Tan, Q.,et al.: Seallms-large language models for southeast asia. In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pp. 294–304 (2024). https://do i.org/10.18653/v1/2024.acl-demos.28

    doi:10.18653/v1/2024.acl-demos.28 2024

  21. [21]

    In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp

    Bui, C.T., Son, N.T., Van Trang, T., Phung, L.V., Huy, P.N., Le, H.A., Van, 20 Q.H., Do, P.N.-T., Truc, V.L.T., Chau, D.T.,et al.: Vmlu benchmarks: A com- prehensive benchmark toolkit for vietnamese llms. In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 11495–11515 (2025). https://doi...

    doi:10.18653/v1/2025.acl-l 2025

  22. [22]

    In: Findings of the Association for Computational Linguistics: NAACL 2024, pp

    Truong, S., Nguyen, D., Nguyen, T.Q., Le, D., Truong, N., Quan, T., Koyejo, S.: Crossing linguistic horizons: Finetuning and comprehensive evaluation of viet- namese large language models. In: Findings of the Association for Computational Linguistics: NAACL 2024, pp. 2849–2900 (2024). https://doi.org/10.18653/v1/ 2024.findings-naacl.182

    doi:10.18653/v1/ 2024

  23. [23]

    Education and information technologies28(12), 15873–15892 (2023) https://doi.org/10.1007/s1 0639-023-11834-1

    Jeon, J., Lee, S.: Large language models in education: A focus on the com- plementary relationship between human teachers and chatgpt. Education and information technologies28(12), 15873–15892 (2023) https://doi.org/10.1007/s1 0639-023-11834-1

    doi:10.1007/s1 2023

  24. [24]

    Education Sciences13(10), 998 (2023) https://doi.org/10.3390/educsci13100998

    Berg, G., Du Plessis, E.: Chatgpt and generative ai: Possibilities for its contri- bution to lesson planning, critical thinking and openness in teacher education. Education Sciences13(10), 998 (2023) https://doi.org/10.3390/educsci13100998

    doi:10.3390/educsci13100998 2023

  25. [25]

    PloS one19(6), 0305337 (2024) https: //doi.org/10.1371/journal.pone.0305337

    Powell, W., Courchesne, S.: Opportunities and risks involved in using chatgpt to create first grade science lesson plans. PloS one19(6), 0305337 (2024) https: //doi.org/10.1371/journal.pone.0305337

    work page doi:10.1371/journal.pone.0305337 2024

  26. [26]

    Performance improvement 42(5), 34–37 (2003)

    Molenda, M.: In search of the elusive addie model. Performance improvement 42(5), 34–37 (2003)

    2003

  27. [27]

    AI and the future of education: Disruptions, dilemmas and directions, 90 (2025)

    Chai, C.S., Wu, J.-Y., Thomas, K.: Keeping the primary goals of education in the al era: What do. AI and the future of education: Disruptions, dilemmas and directions, 90 (2025)

    2025

  28. [28]

    PloS one19(12), 0314688 (2024)

    Luo, F., Tang, C.-H.: Navigating uncertainty: The impact of environmental instability on enterprise digital transformation. PloS one19(12), 0314688 (2024)

    2024

  29. [29]

    In: International Conference on Artificial Intelligence in Education, pp

    Scaria, N., Dharani Chenna, S., Subramani, D.: Automated educational question generation at different bloom’s skill levels using large language models: Strate- gies and evaluation. In: International Conference on Artificial Intelligence in Education, pp. 165–179 (2024). Springer

    2024

  30. [30]

    In: European Conference on Technology Enhanced Learning, pp

    Duong-Trung,N., Wang, X.,Kravčík,M.:Bloomllm: largelanguage models based question generation combining supervised fine-tuning and bloom’s taxonomy. In: European Conference on Technology Enhanced Learning, pp. 93–98 (2024). Springer

    2024

  31. [31]

    Grantee Submission 21 (2019)

    Holstein, K., McLaren, B.M., Aleven, V.: Co-designing a real-time classroom orchestration tool to support teacher-ai complementarity. Grantee Submission 21 (2019)

    2019

  32. [32]

    The Journal of Applied Instructional Design11(4), 69–77 (2022)

    Mayer, R.E.: The future of multimedia learning. The Journal of Applied Instructional Design11(4), 69–77 (2022)

    2022

  33. [33]

    Review of educational research91(2), 204–236 (2021)

    Noetel, M., Griffith, S., Delaney, O., Sanders, T., Parker, P., Pozo Cruz, B., Lonsdale, C.: Video improves learning in higher education: A systematic review. Review of educational research91(2), 204–236 (2021)

    2021

  34. [34]

    In: Fundamentals and Frontiers of Medical Education and Decision-Making, pp

    Schoenherr, J.R., McConnell, M.M.: Human-centered design in health professions education: Informing competency-based education with psychological science. In: Fundamentals and Frontiers of Medical Education and Decision-Making, pp. 3–

  35. [35]

    arXiv e-prints, 2502 (2025)

    Ku, M., Chong, T., Leung, J., Shah, K., Yu, A., Chen, W.: Theoremexplainagent: Towards multimodal explanations for llm theorem understanding. arXiv e-prints, 2502 (2025)

    2025

  36. [36]

    Routledge, ??? (2024)

    2024

  37. [37]

    arXiv:2507.14306 (2025) https://doi.org/10.485 50/arXiv.2507.14306

    Jain, V., Golugula, S., Sathvik, M.S.,et al.: Manimator: Transforming research papers into visual explanations. arXiv:2507.14306 (2025) https://doi.org/10.485 50/arXiv.2507.14306

    doi:10.48550/arxiv.2507.14306 2025

  38. [38]

    Advances in neural information processing systems36, 53728–53741 (2023)

    Rafailov, R., Sharma, A., Mitchell, E., Manning, C.D., Ermon, S., Finn, C.: Direct preference optimization: Your language model is secretly a reward model. Advances in neural information processing systems36, 53728–53741 (2023)

    2023

  39. [39]

    arXiv preprint arXiv:2310.16944 (2023)

    Tunstall, L., Beeching, E., Lambert, N., Rajani, N., Rasul, K., Belkada, Y., Huang, S., Von Werra, L., Fourrier, C., Habib, N., et al.: Zephyr: Direct distillation of lm alignment. arXiv preprint arXiv:2310.16944 (2023)

    Pith/arXiv arXiv 2023

  40. [40]

    In: Findings of the Association for Computational Linguistics: EMNLP 2024, pp

    Sonkar, S., Ni, K., Chaudhary, S., Baraniuk, R.: Pedagogical alignment of large language models. In: Findings of the Association for Computational Linguistics: EMNLP 2024, pp. 13641–13650 (2024)

    2024

  41. [41]

    arXiv preprint arXiv:2412.16429 (2024)

    Team, L., Modi, A., Veerubhotla, A.S., Rysbek, A., Huber, A., Wiltshire, B., Veprek, B., Gillick, D., Kasenberg, D., Ahmed, D., et al.: Learnlm: Improving gemini for learning. arXiv preprint arXiv:2412.16429 (2024)

    Pith/arXiv arXiv 2024

  42. [42]

    arXiv preprint arXiv:2212.08073 (2022)

    Bai, Y., Kadavath, S., Kundu, S., Askell, A., Kernion, J., Jones, A., Chen, A., Goldie, A., Mirhoseini, A., McKinnon, C., et al.: Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073 (2022)

    Pith/arXiv arXiv 2022

  43. [43]

    AI and Ethics, 1–25 (2025)

    Jiao, J., Afroogh, S., Xu, Y., Phillips, C.: Navigating llm ethics: Advancements, challenges, and future directions. AI and Ethics, 1–25 (2025)

    2025

  44. [44]

    arXiv:2408.00118 (2024) https: //doi.org/10.48550/arXiv.2408.00118

    Team, G., Riviere, M., Pathak, S., Sessa, P.G., Hardin, C., Bhupatiraju, S., 22 Hussenot, L., Mesnard, T., Shahriari, B., Ramé, A.,et al.: Gemma 2: Improv- ing open language models at a practical size. arXiv:2408.00118 (2024) https: //doi.org/10.48550/arXiv.2408.00118

    Pith/arXiv arXiv doi:10.48550/arxiv.2408.00118 2024

  45. [45]

    arXiv:2505.09388 (2025) https: //doi.org/10.48550/arXiv.2505.09388

    Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C.,et al.: Qwen3 technical report. arXiv:2505.09388 (2025) https: //doi.org/10.48550/arXiv.2505.09388

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.09388 2025

  46. [46]

    arXiv:2507.06261 (2025) https://doi.org/10.48550/arX iv.2507.06261

    Comanici, G., Bieber, E., Schaekermann, M., Pasupat, I., Sachdeva, N., Dhillon, I., Blistein, M., Ram, O., Zhang, D., Rosen, E.,et al.: Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next genera- tion agentic capabilities. arXiv:2507.06261 (2025) https://doi.org/10.48550/arX iv.2507.06261

    Pith/arXiv arXiv doi:10.48550/arxiv.2507.06261 2025

  47. [47]

    In: eLearn: World Conference on EdTech, pp

    Yang, J.: Innovations in instructional design: integrating chatgpt into addie- based corporate training. In: eLearn: World Conference on EdTech, pp. 108–112 (2024). Association for the Advancement of Computing in Education (AACE). https://www.learntechlib.org/primary/p/224988/

    2024

  48. [48]

    arXiv:2508.19611 (2025) https://doi.org/10.48550/arXiv.2508.19611

    Yao, H., Xu, W., Turnau, J., Kellam, N., Wei, H.: Instructional agents: LLM agents on automated course material generation for teaching faculties. arXiv:2508.19611 (2025) https://doi.org/10.48550/arXiv.2508.19611

    doi:10.48550/arxiv.2508.19611 2025

  49. [49]

    Wiley encyclopedia of clinical trials, 1–3 (2007) https://doi.org/10.1002/9780471462422.eoct979

    Woolson, R.F.: Wilcoxon signed-rank test. Wiley encyclopedia of clinical trials, 1–3 (2007) https://doi.org/10.1002/9780471462422.eoct979

    doi:10.1002/9780471462422.eoct979 2007

  50. [50]

    Computers and Education: Artificial Intelligence2, 100027 (2021)

    Cavalcanti, A.P., Barbosa, A., Carvalho, R., Freitas, F., Tsai, Y.-S., Gašević, D., Mello, R.F.: Automatic feedback in online learning environments: A system- atic literature review. Computers and Education: Artificial Intelligence2, 100027 (2021)

    2021

  51. [51]

    In: INTED2024 Proceedings, pp

    Meyer, L., Dannecker, A.: Smart tools, smarter students: A study on chat- gpt’s effects on learning outcomes and exercise quality in business education. In: INTED2024 Proceedings, pp. 2704–2713 (2024). IATED

    2024

  52. [52]

    Ai Magazine43(2), 239–248 (2022)

    Holstein, K., Aleven, V.: Designing for human–ai complementarity in k-12 education. Ai Magazine43(2), 239–248 (2022)

    2022

  53. [53]

    TPM–Testing, Psychometrics, Methodology in Applied Psychology32(S7 (2025): Posted 10 October), 1322–1335 (2025) 23

    NGUYEN, P.B.T., NGUYEN, T.N.P.: Ethical and responsible ai in educa- tional use of chatgpt. TPM–Testing, Psychometrics, Methodology in Applied Psychology32(S7 (2025): Posted 10 October), 1322–1335 (2025) 23

    2025

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.