International Journal of Education and Humanities (IJEH), 5(4) 2025:751-761
[Link]
E-ISSN: 2798-5768
Enhancing Algorithm Learning with Large Language Models: Design and
Evaluation of AlgoLLM in Higher Education Practice
Shitong Peng1, Yingzhao Lin2, Shoukang Yu3, Jiajun Wu4
Abstract
Algorithm learning remains challenging in computer science education due to its abstract logic, steep
conceptual difficulty, and lack of personalized support in traditional settings. This study presents
AlgoLLM, a modular instructional system built on large language models (LLMs) to support students
through natural language explanations, code-level guidance, and feedback-based refinement. The
system includes four core components: Knowledge Explainer, Exercise Generator, Code Assistant
and Debugger, and Feedback Evaluator. A four-week case study was conducted with 60
undergraduate students, comparing a control group using textbooks and an experimental group using
AlgoLLM. Paired and independent t-tests showed that the experimental group achieved significantly
higher learning gains in post-tests (mean increase of 18.3 percent, Cohen's d = 0.94). Code accuracy
and task efficiency also improved. Pearson correlation revealed a moderate relationship between
LLM interaction frequency and learning gain. Questionnaire feedback indicated high perceived
usefulness, clarity, and satisfaction. These results suggest that LLM-based systems like AlgoLLM
can enhance algorithm comprehension and offer scalable, personalized support in technical
education.
Keywords: Algorithm education, Empirical evaluation, Interactive tutoring systems, Large
language models, Personalized learning.
A. Introduction
Algorithm learning is a foundational yet significantly difficult component of computer
science education, playing a critical role in cultivating students’ abilities in abstract thinking,
logical reasoning, and complex problem-solving (Cormen et al., 2022). Mastering algorithmic
thinking requires not only understanding abstract concepts such as recursion, complexity
analysis, and data structures, but also developing strong problem-solving skills and the ability to
trace logic across multiple steps. However, the abstract nature of algorithms and the demand for
rigorous logical reasoning make algorithmic learning particularly challenging for students—
especially lower-year STEM undergraduates and those from non-technical disciplines (e.g.,
humanities and social sciences). Many learners struggle with understanding problem
decomposition, tracing recursive flows, or interpreting pseudocode, which often leads to
frustration and disengagement. Traditional instructional methods, including lectures, textbooks,
and static problem sets, offer limited personalization and timely feedback. In large classrooms
or online learning environments, teachers face difficulties in adapting explanations to individual
needs or addressing diverse learning paces (Jundan Wang, 2024). Consequently, learners
frequently encounter bottlenecks in understanding complex algorithmic concepts without
sufficient support. Technology-enhanced learning systems with real-time interactivity and visual
1
Department of Mechanical Engineering, Shantou University, Shantou 515063, China. jiajunwu@[Link]
2
Department of Mechanical Engineering, Shantou University, Shantou 515063, China
3
Department of Electrical & Computer Engineering, Northeastern University, Boston 02115, United States
4
Department of Mechanical Engineering, Shantou University, Shantou 515063, China
Peng et al.,
capabilities thus emerge as a critical avenue for addressing these persistent educational barriers,
making algorithmic learning easier for diverse learners.
Recent advances in large language models (LLMs), such as ChatGPT and Claude, offer new
possibilities for addressing these challenges. With capabilities in natural language generation,
contextual understanding, and multi-turn reasoning, LLMs are increasingly prevalent in higher
educational contexts to support content generation, interactive tutoring, and personalized
feedback (Goslen et al., 2025; Michael E. Bernal, 2024). In particular, their ability to generate
stepwise explanations and provide dynamic responses aligns well with the requirements of
algorithm education. Previous studies have demonstrated that AI-assisted systems can
significantly enhance teaching effectiveness and adaptability compared to traditional
approaches. For example, Kreijkes et al. (2025) showed that students using ChatGPT for reading
support scored higher than those using traditional note-taking (d = 0.41). Similarly, Essel et al.
(2022) noted that AI chatbot-assisted students outperformed peers in face-to-face settings (p <
0.05). Holmes et al. (2019) further showed that AI-personalized learning paths increased teacher
adoption of new strategies by 28%. In terms of adaptivity, Tan et al. (2025) reported that AI-
enabled learning systems dynamically adjust content difficulty based on real-time student
performance. Kamalov et al. (2023) demonstrated that adaptive platforms using reinforcement
learning and knowledge tracing improved recommendation accuracy by 8% and enabled real-
time personalized feedback. Latif et al. (2023) further achieved millisecond-level personalized
path optimization using AI ontologies and NLP-generated content with over 95% relevance.
Despite the growing promise of LLMs in education, concerns remain regarding the
reliability of their outputs, including consistency, factual accuracy, and domain relevance
(Sasikala & Ravichandran, 2024). While prior research has explored LLMs in personalized
tutoring and educational dialogue systems, their direct application in algorithm learning remains
limited. Shahzad et al. (2025) highlighted that LLMs may produce plausible yet incorrect
information, posing risks of reinforcing misconceptions among learners. Khan et al. (2025)
reported that although LLMs can identify programming errors, they frequently generate false
positives and redundant suggestions. Jundan Wang (2024), in a recent review, noted that most
existing studies emphasize system design and learner engagement rather than empirically
evaluating how LLMs affect students' understanding of algorithmic logic and problem-solving.
Overall, current work has largely focused on system-level integration or content generation, with
limited attention to their direct instructional impact in structured algorithm education contexts.
To address these gaps, this paper develops an LLM-based learning system designed to
support algorithm education in a higher education context. Through a case study on common
algorithmic topics, we evaluate the system's effectiveness in providing adaptive explanations,
scaffolding problem-solving, and enhancing learners' algorithmic proficiency. Our work
contributes a practical framework for integrating LLMs into technical education, with empirical
findings to guide the development of AI-assisted learning tools.
B. Methods
To investigate the effectiveness of large language models in supporting algorithm
learning, we developed a prototype system called AlgoLLM and conducted a small-scale,
mixed-method case study with undergraduate students in Shantou University, China. This
section presents the system architecture, participant arrangement, experimental procedure,
and evaluation methods used to assess the pedagogical impact.
752
Enhancing Algorithm Learning with Large Language Models…
System design: AlgoLLM architecture
AlgoLLM is an interactive educational assistant that integrates multiple modules to
provide personalized, adaptive, and feedback-rich learning experiences for algorithm
education. The system is implemented in Python and utilizes the GPT-4 model (OpenAI
2024) via OpenAI APIs. LangChain (Chase, 2022) is employed for prompt orchestration,
and the frontend interface is built using Streamlit. Four core functional modules define the
system:
• Knowledge Explainer: Generates step-by-step explanations for algorithmic concepts
using natural language, annotated pseudocode, and analogies to promote conceptual
clarity.
• Exercise Generator: Dynamically creates practice problems tailored to the student’s
level, with progressive difficulty and built-in scaffolding to support learning
trajectories.
• Code Assistant & Debugger: Accepts student-submitted Python code, analyzes logic
and syntax errors, and returns correction suggestions with targeted explanations.
• Feedback & Evaluator: Automatically scores responses, generates concise feedback,
and logs usage metrics for further analysis.
To enhance factual consistency and reduce hallucinations, the system incorporates a
JSON-structured algorithm knowledge base indexed by Pinecone (vector database)
(Pinecone, 2025) and uses semantic retrieval to align LLM outputs with verified content.
Backend request handling is managed via FastAPI, and containerization with Docker ensures
deployability across environments. A complete system workflow and module interaction are
visualized in Fig. 1.
Figure 1 Attached figure in article
753
Peng et al.,
Participants and experimental procedure
The study involved 60 undergraduate computer science students between 18 and 22
years old. All participants had completed coursework in basic data structures but had not
studied the two target algorithms—quicksort and Dijkstra’s algorithm—prior to the
experiment. Participants were randomly assigned to two groups: 30 students in the
experimental group used the AlgoLLM system, while 30 students in the control group used
conventional learning materials. Both groups received equivalent task prompts and content,
ensuring fair comparison.
The experiment lasted four weeks. In Week 1, all participants took a pre-test composed
of multiple-choice conceptual questions and a basic coding task to establish baseline
understanding and implementation ability. During Weeks 2 and 3, the experimental group
used AlgoLLM to complete system-generated exercises and interact with the feedback
modules, while the control group learned independently using PDFs, notes, and static coding
worksheets. Both groups completed the same weekly tasks covering quicksort and Dijkstra’s
algorithm. In Week 4, all students took a post-test structurally parallel to the pre-test.
Additionally, the experimental group completed a Likert-scale questionnaire evaluating the
usefulness, clarity, and satisfaction of the system, and participated in semi-structured
interviews discussing their learning experience and perception of AI-assisted support.
Data collection and evaluation
To assess the learning outcomes and user experience, we collected both quantitative and
qualitative data. Quantitative indicators included pre/post test scores, coding task accuracy,
completion time, and LLM usage metrics such as the number of prompt interactions and
module activation frequency. These were analyzed using paired t-tests (within-group
learning gain) and independent t-tests (between-group differences), with effect sizes
measured by Cohen’s d to indicate practical significance.
Qualitative data were gathered from student interviews and questionnaires. Interviews
focused on participants’ perceptions of clarity, autonomy, and support while interacting with
AlgoLLM. All responses were transcribed and thematically coded using NVivo, with
intercoder agreement exceeding κ = 0.85 to ensure reliability. The questionnaires provided
scaled assessments of students’ satisfaction with system responsiveness, explanation quality,
and overall trust in LLM-generated content. Furthermore, a manual review of AlgoLLM’s
outputs was conducted by two computer science instructors to evaluate factual correctness
and domain alignment of its generated responses.
This mixed-method approach allows for a comprehensive evaluation of AlgoLLM’s
impact on algorithm learning, capturing both measurable performance outcomes and
nuanced learner perspectives in an integrated framework.
C. Results and Discussion
This section presents the observed outcomes and functional assessment of the AlgoLLM
framework. Although a full-scale deployment is pending, we evaluated system readiness and
interaction quality through module simulation, user interface walkthrough, and projected
learning benefits. These results offer insight into how LLM-assisted tools can enhance
algorithm comprehension and engagement.
754
Enhancing Algorithm Learning with Large Language Models…
User interface and interaction design
AlgoLLM features a responsive, web-based interface built with Streamlit, designed to
provide an intuitive and engaging experience for learners, as shown in Fig. 2. The layout
follows a clean two-pane structure: the left sidebar enables module switching and history
navigation, while the main content area hosts a conversational interaction space. The design
facilitates modularity and ease of access across devices, with hover effects, visual feedback,
and graceful error handling mechanisms that improve usability.
Each core module delivers a specific learning function within this interface. The
Knowledge Explainer responds to natural language questions by offering detailed algorithm
breakdowns, including pseudocode, visual examples, and complexity discussions. The
Exercise Generator adapts problem generation to a student’s performance level, creating
scaffolded tasks that progress from basic comprehension to advanced challenge. The Code
Assistant & Debugger accepts Python code input and automatically identifies common
logical or syntactic issues, returning short, targeted guidance to aid correction. The Feedback
& Evaluator assesses submitted code using multiple criteria—correctness, efficiency, and
style—offering not only a numerical score but also brief diagnostics such as “Your
implementation has O(n²) time complexity”. Interaction history is persistently stored and
visible in the sidebar, allowing students to revisit prior queries and maintain continuity in
their learning process.
Figure 2 Interface snapshot of the AlgoLLM
Learning gains and code accuracy analysis
To evaluate the learning effectiveness of AlgoLLM, we applied quantitative statistical
analyses to compare performance between the control and experimental groups. Specifically,
we used paired-sample t-tests to assess within-group improvements from pre-test to post-
test, and independent-sample t-tests to compare post-test scores between the two groups. We
also calculated Cohen’s d to measure the effect size of observed differences. In addition to
mean scores, we report standard deviation (SD) values to reflect the variability of student
performance. A smaller SD indicates more consistent results among participants, while a
larger SD suggests greater dispersion. As shown in Fig. 3, both groups exhibited improved
performance from pre- to post-test, but the experimental group demonstrated significantly
larger gains. The control group’s average score increased from 58.7 (SD = 3.8) to 64.3 (SD
755
Peng et al.,
= 4.9), whereas the experimental group improved from 58.9 (SD = 4.2) to 75.6 (SD = 6.2).
A paired t-test confirmed that the within-group gains were statistically significant for both
groups (p < .01). An independent t-test on post-test scores revealed a significant difference
in favor of the experimental group (t(58) = 8.12, p < .001). The corresponding Cohen’s d =
1.56, indicating a strong effect of the AlgoLLM intervention on learners’ algorithm
understanding. According to conventional benchmarks, a d value above 0.8 represents a
large effect, and values over 1.2 are considered very large, suggesting that AlgoLLM had a
highly meaningful impact.
Figure 3 Learning gains between experimental and control groups
The second metric examined was code implementation accuracy, defined as the ratio of
syntactically correct and semantically valid submissions per student. As shown in Fig. 4, the
experimental group consistently outperformed the control group across all 30 participants.
While the control group’s accuracy ranged primarily between 0.65–0.78, the experimental
group maintained a higher accuracy cluster in the 0.85–0.95 range. This suggests that
AlgoLLM’s code assistant and real-time feedback modules contributed substantially to
reducing typical programming errors such as off-by-one mistakes and incorrect loop logic.
Figure 4 Comparison of code accuracy by group
756
Enhancing Algorithm Learning with Large Language Models…
To investigate whether the frequency of interaction with AlgoLLM was associated with
greater learning benefits, we analyzed the relationship between the number of interaction
rounds and the individual score gains in the experimental group. As shown in Fig. 5, the
scatter plot reveals a modest positive trend, with the linear regression line indicating that
higher interaction frequency generally correlates with greater post-test improvement. While
individual variance exists, the upward slope of the trend line suggests that students who
engaged more actively with the system tended to achieve higher score gains. This finding
supports the hypothesis that repeated exposure to step-by-step explanations, practice
feedback, and code correction contributes cumulatively to conceptual consolidation and
problem-solving accuracy. The result aligns with previous studies emphasizing the role of
iterative interaction in adaptive tutoring systems.
Figure 5 Relationship between LLM interaction frequency and learning improvement
Framework adaptivity and feedback quality
AlgoLLM was purposefully built to adapt to diverse learner needs. It dynamically
adjusts exercise difficulty and explanation depth based on user inputs, allowing for more
personalized content delivery. Early feedback from test users indicates a high satisfaction
level with the quality and clarity of LLM-generated responses, averaging 4.6 out of 5 in post-
session surveys. The system reliably identifies common coding issues (e.g., off-by-one
errors and missing base cases) and offers feedback that is concise yet actionable.
To gain deeper insights into learners’ subjective experiences, we collected both
quantitative and qualitative feedback from the experimental group (n = 30) following the
four-week intervention. A post-study questionnaire was administered with three Likert-scale
items designed to measure students’ perceptions of the system’s usefulness, clarity of
explanations, and overall satisfaction.
The summary statistics of the responses are shown in Table 1. On average, students
rated all three aspects highly, with usefulness (Mean, M = 4.55) and satisfaction (M = 4.63)
being the most positively evaluated dimensions. SD values across items were moderate,
suggesting consistent agreement among participants. In particular, students frequently
mentioned that the system’s step-by-step explanations helped them better grasp abstract
topics such as recursion and graph traversal.
757
Peng et al.,
Table 1 Summary of students’ Likert-scale ratings (1–5 scale)
Metric Mean score Standard Interpretation
deviation
Usefulness 4.55 0.62 Rated highly for aiding algorithm learning
Clarity 4.38 0.71 Explanations perceived as mostly clear
Satisfaction 4.63 0.49 Strong overall approval from participants
In addition to the survey, 12 students participated in semi-structured interviews. Thematic
analysis revealed several recurring insights:
• Clarity and Confidence Building: Many students reported that the system’s breakdown of
complex algorithms (e.g., recursion) helped them feel less intimidated and more confident.
One student stated, “It felt like the AI could predict where I’d get confused, and it explained
things right before I asked.”
• Debugging Support: Several students praised the Code Assistant module for identifying
logical or syntactic issues in real time. “Instead of searching forums for hours, I fixed bugs
in minutes,” one participant said.
• Adaptive Practice: Students appreciated the dynamic difficulty of exercises, with some
calling it “the most personalized way I’ve learned programming so far.”
• Suggestions: While generally positive, some students recommended richer visual content
and better navigation of historical chats.
These results suggest that AlgoLLM not only delivers functional support but also provides
an engaging and confidence-enhancing learning experience. The combination of adaptive
scaffolding and conversational interaction appears especially valuable in algorithm education.
The observed outputs of AlgoLLM, as reflected in the test score gains (Fig. 3), code
accuracy improvements (Fig. 4), and usage–performance correlation (Fig. 5), suggest that LLM-
based systems can be a promising medium for enhancing algorithm education, particularly in
environments where individualized feedback and conceptual scaffolding are limited. By offering
a modular combination of stepwise explanation, interactive code diagnostics, and adaptive
practice generation, AlgoLLM enables learners to break down complex problems, reduce
misconceptions, and incrementally build problem-solving fluency. Compared to traditional static
materials or delayed-response help systems (e.g., forums), AlgoLLM provides an “always-
available” conversational tutor that is both responsive and context-aware. Its design reflects an
understanding that algorithm learning is not purely syntactic but involves deeply logical and
sequential reasoning. While prior studies have applied LLMs to writing or language learning
contexts, few have tailored them to support algorithmic reasoning and code-level diagnostics.
Our integration of symbolic knowledge (via concept graphs) and programmatic analysis enables
the system to bridge gaps between natural language understanding and programming logic.
Notably, Fig. 5 reveals a modest positive correlation between interaction frequency and learning
gains. This suggests that iterative engagement with LLM feedback—particularly through the
Code Assistant and Knowledge Explainer modules—may reinforce understanding through
active retrieval, feedback loops, and immediate clarification of misconceptions. Such findings
align with theories of cognitive apprenticeship and retrieval-based learning.
Nevertheless, several limitations must be acknowledged. While LLM outputs were
generally coherent, their quality depended heavily on prompt phrasing, and inconsistencies were
observed in edge cases. To mitigate hallucinations, our system relies on a curated concept base;
758
Enhancing Algorithm Learning with Large Language Models…
however, real-time retrieval and scalable factual grounding remain ongoing challenges.
Moreover, although the interface encourages rich interaction, there is a risk of over-reliance on
AI-generated explanations. Students may accept outputs uncritically, bypassing deeper cognitive
engagement. Embedding confidence scores or uncertainty indicators may promote
metacognitive awareness, helping learners reflect on when to trust or challenge the system. The
broader implications of integrating LLMs into technical education also deserve reflection. As
these systems scale, questions around instructor acceptance, ethical transparency, and alignment
with curricular goals become critical. We argue that AlgoLLM and similar tools should not be
viewed as replacements, but as pedagogically grounded assistants co-designed with educators.
Interpretability and feedback traceability features may help strengthen teacher trust and student
accountability.
Looking ahead, future research should evaluate AlgoLLM in real-world classrooms,
exploring its role as a supplemental tutor, formative assessment tool, or flipped-classroom
assistant. Longitudinal studies across diverse learner populations could illuminate how sustained
use influences algorithmic thinking, motivation, and skill transfer. Furthermore, expanding
input/output modalities (e.g., sketch recognition, voice, emotional cues) may broaden
accessibility and deepen personalization. As LLMs become faster and more cost-efficient, they
hold the potential to democratize access to high-quality algorithm education, especially in under-
resourced settings.
Overall, the findings reinforce the idea that LLMs, when embedded in pedagogically
informed frameworks, can meaningfully support technical skill acquisition. With continued
improvements in accuracy, adaptivity, and transparency, systems like AlgoLLM may help shape
the next generation of human-AI collaborative learning environments.
D. Conclusion
This paper presents AlgoLLM, a large language model-based instructional framework
designed to assist students in learning algorithmic concepts through interactive explanation,
adaptive practice, and real-time feedback. By integrating modules such as Knowledge Explainer,
Code Assistant, and Evaluator into a unified web interface, the system offers personalized
learning support that aligns with individual cognitive levels and problem-solving needs. Our
preliminary walkthrough and simulated interaction results suggest that LLM-enhanced
environments can improve conceptual understanding, reduce time spent on debugging, and
promote more iterative and reflective learning behaviors. Compared with traditional teaching
methods, AlgoLLM facilitates dynamic, context-aware assistance that is particularly valuable in
abstract domains like recursion, sorting, and graph algorithms.
However, several challenges remain. These include ensuring the factual consistency of
generated responses, avoiding over-dependence on AI guidance, and integrating such systems
meaningfully into real classroom settings. Future research should focus on large-scale
deployment, long-term outcome measurement, and teacher–AI co-orchestration models.
In addition to measurable performance improvements, student feedback revealed high levels
of perceived usefulness and satisfaction with AlgoLLM. The consistently strong Likert-scale
ratings (Table 3) suggest that learners not only benefited cognitively but also experienced
positive affective engagement. Interview responses indicated that many students found the
system to be clearer and more responsive than traditional materials, particularly in handling
challenging topics such as recursion and graph traversal. These findings underscore the value of
aligning technical features (e.g., code feedback, adaptive practice) with user experience design.
While prior research has emphasized learning outcomes, our results show that student motivation
759
Peng et al.,
and confidence may also be significantly enhanced by well-integrated LLM systems. This
highlights the dual pedagogical role of AlgoLLM: as both a performance-enhancing tutor and a
confidence-building learning companion.
Despite these promising findings, the system has several limitations. AlgoLLM is intended
as an assistant rather than a full replacement for human instruction. It still struggles with complex
reasoning chains, ambiguity, and high-context queries. Its adaptivity is limited to short-term
interactions and does not fully model long-term learner knowledge states. There is also a risk of
student over-reliance on AI responses, which may reduce critical thinking and independent
problem-solving skills.
Future work may involve deploying AlgoLLM in distributed, classroom-scale settings to
test scalability and robustness. The framework could be extended with automated data analysis
modules that detect common learning bottlenecks, misconceptions, and coding behavior
patterns. These insights would enable faster and more accurate feedback delivery. Integrating
multimodal input (e.g., diagrams, voice, or sketch recognition-aware interaction) may also
improve engagement and accessibility. By aligning instructional intelligence with cognitive
modeling and human-centered design, AlgoLLM has the potential to evolve into a next-
generation, equitable platform for algorithm education
Acknowledgment
The authors would like to thank the 60 undergraduate students from the Intelligent
Manufacturing major at Shantou University for their active participation in this study. Their
engagement and valuable feedback were essential to the development and evaluation of the
AlgoLLM system. We also appreciate the financial support from Shantou University Education
and Teaching Reform Project's 'Exploration and Practice of Generative AI-Enabled Personalized
and Project-Based Teaching Strategies in the Machine Learning Course'
References
Chase, H. (2022, October). LangChain. [Link]
Cormen, T. H., Leiserson, C. E., Rivest, R. L., & Stein, C. (2022). Introduction to Algorithms, fourth
edition. MIT Press.
Essel, H. B., Vlachopoulos, D., Tachie-Menson, A., Johnson, E. E., & Baah, P. K. (2022). The impact
of a virtual teaching assistant (chatbot) on students’ learning in Ghanaian higher education.
International Journal of Educational Technology in Higher Education, 19(1).
[Link]
Goslen, A., Kim, Y. J., Rowe, J., & Lester, J. (2025). LLM-Based Student Plan Generation for
Adaptive Scaffolding in Game-Based Learning Environments. International Journal of Artificial
Intelligence in Education, 35(2), 533–558. [Link]
Holmes, W. (2019). Artificial intelligence in education: Promises and implications for teaching and
learning. Center for Curriculum Redesign.
Jundan Wang. (2024). Application and Exploration of Cloud Technology in University Classroom
Teaching—Take Cloud Classroom as an Example. International Journal of Interdisciplinary
Research and Innovations, 12,(2348–1226), 55–62.
[Link]
760
Enhancing Algorithm Learning with Large Language Models…
Kamalov, F., Santandreu Calonge, D., & Gurrib, I. (2023). New Era of Artificial Intelligence in
Education: Towards a Sustainable Multifaceted Revolution. Sustainability, 15(16), 12451.
[Link]
Khan, M., Akbar, M. A., & Kasurinen, J. (2025). Integrating LLMs in Software Engineering
Education: Motivators, Demotivators, and a Roadmap Towards a Framework for Finnish Higher
Education Institutes (No. arXiv:2503.22238). arXiv.
[Link]
Kreijkes, P., Kewenig, V., Kuvalja, M., Lee, M., Vitello, S., Hofman, J., Sellen, A., Rintel, S.,
Goldstein, D. G., Rothschild, D. M., Tankelevitch, L., & Oates, T. (2025). Effects of LLM Use
and Note-Taking On Reading Comprehension and Memory: A Randomised Experiment in
Secondary Schools. Elsevier BV. [Link]
Latif, M., Abbasi, F. L., Ammar, M., Mehmood, B., & Ali, S. (2023). Transforming Personalized
Education through AI- Enhanced Ontology Modelling in Dynamic Adaptive Learning Systems.
International Journal on Recent and Innovation Trends in Computing and Communication,
12(2).
Michael E. Bernal. (2024). Revolutionizing eLearning Assessments: The Role of GPT in Crafting
Dynamic Content and Feedback. Journal of Artificial Intelligence and Technology.
[Link]
OpenAI, Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D.,
Altenschmidt, J., Altman, S., Anadkat, S., Avila, R., Babuschkin, I., Balaji, S., Balcom, V.,
Baltescu, P., Bao, H., Bavarian, M., Belgum, J., … Zoph, B. (2024). GPT-4 Technical Report
(No. arXiv:2303.08774). arXiv. [Link]
Pinecone. (2025). The vector database to build knowledgeable AI. [Link]
Sasikala, P., & Ravichandran, R. (2024). Study on the Impact of Artificial Intelligence on Student
Learning Outcomes. Journal of Digital Learning and Education, 4, 145–155.
[Link]
Shahzad, T., Mazhar, T., Tariq, M., Ahmad, W., & Ouahada, K. (2025). A comprehensive review of
large language models: Issues and solutions in learning environments. Discover Sustainability,
6. [Link]
Tan, K., Yao, J., Pang, T., Fan, C., & Song, Y. (2025). ELF: Educational LLM Framework of
Improving and Evaluating AI Generated Content for Classroom Teaching. Journal of Data and
Information Quality. [Link]
761