Source-linked AI summary
Large Language Models for Education: A Survey and Outlook
Shen Wang, Tianlong Xu, Hang Li, Chaoli Zhang, Joleen Liang, Jiliang Tang, Philip S. Yu, Qingsong Wen
TL;DR
Education research has lacked a systematic, technology-centered synthesis of LLM applications, datasets, benchmarks, risks, and future directions. This survey develops a multifaceted taxonomy and reviews these resources and challenges, reporting broad progress alongside concerns about fairness, accountability, and overreliance.
Problem
Existing literature has not systematically summarized LLMs for education from a technological perspective, despite extensive application-focused work.
Method
The survey organizes LLM technologies across student and teacher assistance, adaptive learning, and other educational scenarios, while reviewing datasets, benchmarks, challenges, and future directions.
Results
The survey reports broad progress in educational LLM applications, including question solving, error correction, and teacher-oriented material creation.
Takeaways & Limitations
The survey provides educators, researchers, and policymakers with a comprehensive technological picture for developing and assessing LLM applications in education.
Takeaways & Limitations
Educational deployment must address unequal access, transparency and accountability concerns, and the risk that overreliance harms students’ critical thinking, writing, and creativity.
Abstract
from arXiv · showhide
The advent of Large Language Models (LLMs) has brought in a new era of possibilities in the realm of education. This survey paper summarizes the various technologies of LLMs in educational settings from multifaceted perspectives, encompassing student and teacher assistance, adaptive learning, and commercial tools. We systematically review the technological advancements in each perspective, organize related datasets and benchmarks, and identify the risks and challenges associated with deploying LLMs in education. Furthermore, we outline future research opportunities, highlighting the potential promising directions. Our survey aims to provide a comprehensive technological picture for educators, researchers, and policymakers to harness the power of LLMs to revolutionize educational practices and foster a more effective personalized learning environment.
1 INTRODUCTION
LLMs have achieved strong performance across educational scenarios and can assist with writing and reading. Existing literature covers applications and visions, but this survey addresses the lack of a systematic technology-centered review.
- Motivation: LLMs achieve state-of-the-art performance in multiple educational scenarios and student-level performance on standardized tests across mathematics subjects.The reported test settings include multiple-choice and free-response problems.
- Motivation: Empirical studies identify LLMs as writing and reading assistants capable of logically consistent answers across disciplines.One study describes a balance between answer depth and breadth.
- Related Work: Prior surveys and exploratory studies analyze pedagogical responses, user perspectives, and published literature on LLM applications in education.These works include analyses of engineering education, interviews, online posts, user logs, and structured literature tables.
- Contribution: The survey argues that existing literature has not systematically summarized LLMs for education from a technological perspective.It therefore presents a technology-centric taxonomy, publicly available datasets and benchmarks, current challenges, and future opportunities.
2.1 Overview
The paper organizes education applications according to users’ roles and usage scenarios, then summarizes LLM appearances and benefits through a taxonomy.
- 2.1 Overview: Education applications are categorized by users’ roles and usage scenarios.The taxonomy is used to organize LLM applications across education settings.
- 2.1 Overview: The survey summarizes LLM applications and discusses their benefits compared with original methods.A primary summary is presented through the taxonomy illustrated in Figure 1.
2.2 Study Assisting
LLMs support students through question solving, error correction, and guidance, offering timely assistance across subjects. However, generated explanations and hints can be less effective than teacher-written or human-tutor-generated support.
- Study Assisting: Timely learning support is important for student engagement and learning efficiency during independent study.Prior fixed-form-response algorithms often generalize poorly in real-world scenarios.
- Study Assisting: LLM-based educational support provides real-time help by solving challenging questions, correcting errors, and offering explanations or hints.These capabilities address limitations associated with fixed-form responses.
- Question Solving: LLMs have progressed in question solving across subjects, with techniques such as Chain-of-Thought proposed for complicated questions.The survey describes resulting progress and timely access to answers for students’ blocking questions.
- Error Correction: Error-correction studies report strong potential, including perfect resolution of some simple spelling errors using prompt strategies.The cited strategies include zero-shot, zero-shot-CoT, few-shot, and few-shot-CoT.
- Confusion Helper: Synthetic explanations and LLM-generated algebra hints can be less effective than teacher-written explanations and human-tutor-generated hints.Reported issues include unfamiliar terms, overly general advice, and lower learning effectiveness.
2.3 Teach Assisting
LLMs assist teachers with routine educational work, including question generation, grading, and material creation. The reviewed approaches use controllable generation, diversity controls, and human-in-the-loop processes to improve outputs.
- Teach Assisting: Teach-assisting models aim to reduce instructors’ routine workloads so they can focus on in-class instruction.The passage contrasts these tasks with instruction that existing machine-learning models cannot replace.
- Question Generation: Question-generation approaches fine-tune LLMs with reading materials and textbook exercises, then use controllable generation to produce coherent passages.Aligned prompts can cover varied question types and difficulty levels in a balanced manner.
- Question Generation: Math word-problem generation uses implicit diversity controls toward the equation of the question to produce diverse, high-quality questions.The cited approach targets diversity while preserving question quality.
- Material Creation: LLM integration with a human-in-the-loop process is used in asynchronous course creation to improve the accuracy and clarity of generated educational content.The application concerns adult learning, training, and upskilling.
- Material Creation: Generated work examples receive positive feedback from students after GPT-4 is optimized with one-shot learning.The cited result concerns educational material creation.
2.4 Adaptive Learning
Adaptive learning applications using LLMs are organized around knowledge tracing and personalized content generation. These approaches use LLMs to enrich student and question information and to generate learning paths informed by students’ knowledge mastery.
- Adaptive learning: Adaptive learning work is classified into knowledge tracing and content personalization.Knowledge tracing estimates students’ knowledge mastery from response correctness, while content personalization provides customized learning content based on personalized factors.
- Knowledge tracing: Knowledge-tracing systems use LLMs to generate auxiliary information for questions and student records.One approach extracts knowledge keywords for questions in a student-question response graph, with particular advantages in cold-start scenarios.
- Content personalization: Personalized-content systems generate dynamic learning paths based on recent knowledge diagnoses and incorporate knowledge-concept structures.Generation can select subsequent learning objects according to whether a student has mastered a topic.
2.5 Education Toolkit
Commercial LLM-powered education tools span chatbots, content creation, teaching aides, quiz generators, and collaboration tools. They support personalized assistance, educational-material preparation, feedback, assessment generation, and discussion or video-based learning activities.
- Toolkit overview: Commercial LLM education tools are categorized as Chatbot, Content Creation, Teaching Aide, Quiz Generator, and Collaboration Tool.This taxonomy organizes industry tools by their educational usage scenarios.
- Chatbot: LLM chatbots provide personalized feedback and support, continuous availability, and scalable handling of learner queries.They can also automate repetitive tasks such as quiz grading and basic feedback.
- Content Creation: Content-creation tools generate interactive slide decks, leveled resources, customized materials, lesson plans, assessments, and other educational content.Examples include interactive presentations, adapted reading resources, automated planning, and interactive courses or videos.
- Teaching Aide: Teaching-aide tools support personalized feedback, writing assistance, neurodivergent users, and conversational interaction with PDF documents.Examples include gotFeedback, Grammarly, Goblin Tools, and ChatPDF.
- Quiz Generator and Collaboration: Quiz-generation tools create questions, learning objectives, aligned multiple-choice items, hints, and feedback from supplied educational materials.Some tools also modify existing quizzes or generate discussion prompts and higher-order thinking questions from videos or articles.
3 DATASET AND BENCHMARK
Educational datasets and benchmarks support text-rich downstream tasks and cover diverse aspects of the educational process. Question solving is especially prevalent across education and natural language processing, with many datasets focused on mathematics.
- Dataset and benchmark scope: LLM-based educational datasets and benchmarks are constructed for text-rich downstream tasks such as question solving and error correction.These tasks leverage LLM knowledge and linguistic understanding for comprehension, generation, or text transformation.
- Dataset coverage: Educational datasets vary across student performance, text and resource databases, online learning, language learning, games, demographics, socioeconomic factors, LMS data, and special education.Their scope reflects multiple aspects of the educational process.
- Question solving: Question-solving datasets constitute a significant share because question solving is prevalent in both education and NLP.Many of these datasets are constructed for mathematics question-solving tasks.
4 RISKS AND POTENTIAL CHALLENGES
Deploying LLMs in education raises concerns about fairness, reliability, transparency, privacy, and student overreliance. The survey emphasizes that educational use requires caution because model limitations can affect access, trust, academic integrity, data control, and independent thinking.
- Fairness and Inclusiveness: LLM training data can reproduce dominant group representations and social stereotypes, creating bias in educational applications.Fairness concerns include demographic bias and counterfactual concerns.
- Scope boundary: The survey omits some educational applications because publicly accessible datasets were unavailable.The stated examples include confusion helpers, material creation, knowledge tracking, and content personalization.
- Fairness and Inclusiveness: Some demographic groups may lack equal access to educational models of comparable quality, especially when their languages are underrepresented in training data.The survey also describes prompting, few-shot learning, retrieval augmentation, and bias indices as mitigation or evaluation approaches.
- Reliability and Safety: LLMs face reliability issues including hallucinations, toxic outputs, inconsistent responses, misinformation, and temporal misalignment.These problems are particularly significant in education, where generated content can confuse or mislead students.
- Transparency and Accountability: Black-box operation creates transparency and accountability concerns, alongside challenges involving plagiarism, cheating, inaccurate reporting, assessment, and academic integrity.Proposed responses include new assessment frameworks, policy reconsideration, machine-text detectors, and citation during training.
- Privacy and Security: Education-sector privacy and security require heightened scrutiny because learner awareness and acceptance of tracking and profiling remain underexplored.The text emphasizes user control over personal data, including the ability to delete interactions.
- Overreliance: Students’ blind reliance on LLMs may undermine independent thinking and writing-skill development.The survey discusses moderated approaches and argues that LLMs should augment rather than substitute human thinking capacities.
5 FUTURE DIRECTIONS
The paper identifies future directions spanning pedagogically aligned, multi-agent, multilingual and multimodal, edge-deployed, specialized, and ethically governed LLMs for education. These directions aim to improve alignment, personalization, efficiency, accessibility, subject relevance, and responsible deployment.
- 5.1 Pedagogical Interest Aligned LLMs: Real-world instruction remains challenging because high-quality education requires multidisciplinary knowledge and must satisfy administrative constraints.The paper proposes retrieval-augmented generation and real-world pedagogical instruction data for better alignment with instructors’ behavior.
- 5.2 Multi-Agent-based LLMs: Multi-agent LLM systems are presented as a promising fit for educational problems involving multistep processing logic.The paper connects this direction to collaboration frameworks in which multiple LLMs solve complicated tasks through conversation.
- 5.3 Multimodal and Multilingual Supports: Multilingual and multimodal LLMs could support native-language access and richer analysis of student engagement, comprehension, and learning styles.The paper highlights integrating varied data sources for more nuanced, personalized learning experiences and improving equity and inclusion through culturally aware multilingual models.
- 5.4 Edge Computing and Efficiency: Edge computing could improve educational technology efficiency, accessibility, offline availability, privacy, and security by processing data closer to end users.The paper proposes lightweight models that preserve performance while reducing computational requirements, especially where internet connectivity is limited.
- 5.5 Efficient Training of Specialized Models: Specialized LLMs tailored to educational subjects could provide more accurate, relevant, and cost-effective support for students and educators.Efficient training remains challenging and requires advances in data collection, model architecture, and training methodologies.
- 5.6 Ethical and Privacy Considerations: Ethical LLM deployment in education requires attention to data security, student privacy, bias mitigation, transparency, and inclusive model development.The paper calls for frameworks and guidelines governing responsible use in educational settings.
6 CONCLUSION
The survey reviews LLM applications across student and teacher assistance, adaptive learning, and miscellaneous tools, while organizing datasets, benchmarks, challenges, and future directions. It aims to support innovative work on LLMs for education.
- 6 CONCLUSION: The survey provides a multifaceted review of LLMs applied to student assistance, teacher assistance, adaptive learning, and miscellaneous educational tools.It also summarizes related datasets and benchmarks, current challenges, and future research directions.
- 6 CONCLUSION: The survey aims to facilitate and inspire more innovative work within LLMs for education.
A APPENDIX
The appendix identifies Table 1 as a summary of existing datasets and benchmarks for evaluating LLMs on education applications. The table provides a reference for the datasets and benchmarks used in this area.
- A APPENDIX: Table 1 summarizes commonly used publicly available datasets and benchmarks for evaluating LLMs on education applications.