LLM Agents for Education: Advances and Applications

1 Introduction

• Traditional educational data mining struggles with shallow contextual understanding, limited interaction, and generating adaptive personalized materials, despite progress in analyzing student behavior and knowledge states.LLM agents’ language understanding and task automation, supported by memory, tool use, and planning, are presented as ways to address these limitations.Source: §1 · p. 1
• Personalization, explainability, and multi-agent communication complement memory, tool use, and planning by supporting tailored instruction, interpretable decisions, and collaborative reasoning.The survey associates these features with improving understanding and engagement while streamlining educational workflows.Source: §1 · p. 1
• The survey’s task-centric taxonomy organizes educational agents into Teaching Assistance and Student Support roles across six core educational tasks.Teaching Assistance covers classroom simulation, feedback comment generation, and curriculum design; Student Support covers adaptive learning, knowledge tracing, and error correction and detection.Source: §1 · p. 2
• The survey highlights ethical issues, hallucination and overreliance, and integration with existing educational ecosystems as critical deployment challenges.It also compiles datasets, benchmarks, and evaluation protocols to support future research on educational LLM-agent systems.Source: §1 · p. 2
• Appendix A reviews domain-specific agents in science learning, language learning, and professional development, including their challenges, advances, benchmarks, and datasets.The overview supplements the survey’s main focus on technical foundations and general-purpose pedagogical capabilities.Source: §1 · p. 2

Memory

• Memory combines long-term knowledge with short-term interaction data, while summarization and retrieval help preserve relevant context for tracking student progress and personalizing responses.Filtering noisy interaction data is important to maintain memory quality.Source: §2 · p. 2

Tool Use

• External tools such as search engines, databases, and APIs help agents address knowledge-cutoff and calculation limits while supporting access to current information and specialized capabilities.Source: §2 · p. 2

Planning

• Planning breaks complex learning goals into steps, adapts learning paths to student progress, and can support dynamic curricula co-created with students.For longer-term tasks, agents can act as metacontrollers that stay aligned with teaching goals through approaches such as “Pedagogical Steering.”Source: §2 · pp. 2–3

Personalization

• Personalization lets agents adapt to individual learning needs as tutors, teaching assistants, simulated peers, or tools for recommending educational content.Source: §2 · p. 3

Explainability and Multi-Agent Communication

• Clear, step-by-step explanations support trust and student understanding, especially in STEM, while multi-agent coordination can foster critical thinking and diverse perspectives in collaborative learning.Agents can also simulate teamwork and professional settings when guided by teaching frameworks for organization and evaluation.Source: §2 · p. 3

Capability-to-task mapping

• Table 1 maps primary capabilities to educational tasks, while noting that studies explore additional capabilities, such as multi-agent communication to improve feedback correctness (Guo et al. (2024c)).The table omits secondary capabilities to clarify each task’s primary focus.Source: §2 · p. 3

3.1 Classroom Simulation

• Classroom simulations model student behavior across diverse personas and learning histories, enabling educators to test strategies tailored to different student profiles.Studies report behavior predictions aligned with teachers’ expectations (Xu et al., 2024a; Jin et al., 2024b; Li et al., 2025a); CGMI uses memory, reflection, and planning, while Classroom Simulacra adds iterative reflection. Simulations can also help refine teaching plans and instructional strategies.Source: §3.1 · p. 4

3.2 Feedback Comment Generation

• Feedback agents use multi-agent review, reinforcement learning, and student-teacher roleplay to generate or improve comments on student work.For example, Guo et al. (2024c) use a second agent to refine initial feedback; PROF (Nair et al., 2024) rewards feedback quality, and SEFL (Zhang et al., 2025a) generates training data through roleplay. Agents may still give inaccurate feedback on complex programming and professional review tasks (Lou et al., 2024).Source: §3.2 · p. 5

3.3 Curriculum Design

• Curriculum agents recommend existing resources through retrieval or generate content adapted to students’ styles, knowledge gaps, and interests.Explaining recommendations can support trust and informed choices; knowledge graphs can guide pathways toward curated sources and reduce misinformation risk (Abu-Rasheed et al., 2024a,b). Future directions include integration with adaptive learning systems, hybrid retrieval-generation methods, and multimodal resources.Source: §3.3 · p. 5

4.1 Adaptive Learning

• GenAL uses external tools to build student profiles, while EduAgent models four cognitive patterns and Chen et al. (2024c) combine interaction, reflection, reaction, and meta-agent control.GenAL dynamically updates memory and assigns tasks suited to learners’ knowledge levels; EduAgent profiles gaze, motor behavior, cognitive state, and post-course assessments.Source: §4.1 · p. 6
• Adaptive state models track cognitive, affective, preference, and personality traits; Wang et al. (2025b) use five specialized agents, and OATutor supports modular tutoring.Cognitive states capture proficiency, comprehension, and misconceptions; affective and psychological modeling can shape feedback tone, pacing, and instructional format.Source: §4.1 · p. 6

4.2 Knowledge Tracing

• Knowledge tracing estimates evolving understanding and future performance; Yang et al. (2024b) use administrator, judger, and critic agents, while Xu et al. (2024a) and Scarlatos et al. (2025) explore persona-based and dialogue-driven approaches.Judgers discuss assessments of the student’s cognitive state, and a critic checks whether assessment criteria are met; dialogue can probe conceptual boundaries and refine knowledge estimates.Source: §4.2 · pp. 6–7

4.3 Error Correction and Detection

• LLM agents detect errors across writing, programming, and math, using learner state and adaptive inference; Xu et al. (2024b) analyze drafts, while CoT Rerailer (Wan et al., 2024) detects and rerails math-solving errors.Xu et al. convert handwritten or digital drafts into natural language and provide indirect feedback that guides self-correction; MathCCS (Zhang et al., 2025d) and ErrorRadar (Yan et al., 2024c) support multimodal error analysis.Source: §4.3 · p. 7

5.1 Privacy, Bias and Fairness

• Sensitive data create privacy risks, while training data can reinforce stereotypes and disparities; proposed directions include unlearning, automated fairness checks, and culturally adaptive multilingual agents.Unlearning aims to retain useful knowledge while selectively forgetting sensitive user data.Source: §5.1 · p. 7

5.2 Hallucination and Overreliance

• Hallucinations can mislead learners and overreliance can hinder skill acquisition; safeguards include self-reflecting tutors (Renze and Guven, 2024), teacher review, and pedagogically aligned agents.Source: §5.2 · pp. 7–8

5.3 Integration with Existing Educational Ecosystems

• Integration is constrained by absent deployment frameworks, limited PBL guidance, and unequal infrastructure; research should standardize deployment and measure learning gains, trust, and engagement.FOKE combines foundation models, knowledge graphs, and prompt engineering (Hu and Wang, 2024), while AI-VERDE provides unified access, privacy, and budget controls (Mithun et al., 2025). Directions include adaptive and multimodal feedback, affordable cloud or decentralized tutors, and education-oriented benchmarks and datasets.Source: §5.3 · p. 8

5.4 Multimodal LLM Agents for Education

• Multimodal agents struggle with modality integration, cultural and contextual alignment, and low-latency reasoning; proposed remedies balance modality fusion, personalize to context, and optimize latency.Suggested methods include contrastive pretraining and modality-specific adapters, localized knowledge graphs with human feedback, and lightweight distilled models with selective modality routing.Source: §5.4 · pp. 8–9

• The task-centric taxonomy groups educational LLM agents into Teaching Assistance and Student Support, with capabilities including memory augmentation, tool use, planning, and personalization.Source: §6 · p. 9
• Reliable and ethical deployment requires addressing ethical issues, hallucination and overreliance, and integration with existing educational ecosystems.Source: §6 · p. 9
• The survey compiles critical datasets, benchmarks, and evaluation methodologies to support continued progress in the field.Source: §6 · p. 9
• LLM agents have potential to transform personalized learning, intelligent tutoring, and pedagogical automation, but realizing their full potential requires technical rigor and thoughtful system design.Source: §6 · p. 9

Limitations

• The review acknowledges that some recent advances may have been missed, while aiming to include foundational and representative works.Source: Limitations · p. 9
• The literature search used Google Scholar and covered publications from 2020 to May 2025.Source: Limitations · p. 9
• Search terms included “LLM agents for education,” classroom simulation, adaptive learning, and knowledge tracing.Source: Limitations · p. 9
• The review included studies on LLM agent-based methods but excluded studies focused solely on LLM-based approaches.Source: Limitations · p. 9

A Domain-Specific Educational Agents

• Domain-specific applications covered science learning, language learning, and professional development.Source: A Domain-Specific Educational Agents · p. 23
• The domain-specific review examines algorithmic frameworks, agentic designs, datasets, and benchmarks, and presents a systematic taxonomy in Figure 3.Source: A Domain-Specific Educational Agents · p. 23

A.1 Agent for Science Learning

• Science-learning agents use personalized, interactive experiences to help students acquire and apply scientific knowledge, with tailored feedback, conceptual support, and active engagement.The broader survey maps their applications across mathematics, physics, chemistry, biology, and general scientific discovery.Source: §A.1 · p. 23

A.1.1 Mathematics

• Mathematics agents support complex problem solving and abstract concepts, including TORA’s natural-language and program-based reasoning and MathAgent’s multimodal error detection and visual-to-formal conversion.MathAgent targets real-world K–12 scenarios, such as converting plane-geometry images into formal expressions.Source: §A.1.1 · p. 23
• MathChat combines a chat-based LLM agent with a tool-based user agent for conversational mathematical problem solving.Wu et al. (2023b) introduce the system.Source: §A.1.1 · p. 23
• Xiong et al. (2024) formulate RLHF for tool-integrated mathematical agents as a Markov decision process, while MACM uses interactive Thinker, Judge, and Executor agents for complex deduction.The RLHF formulation contrasts with the typical contextual-bandit approach; MACM (Lei et al., 2025) addresses LLM limitations in complex logical deduction.Source: §A.1.1 · p. 23

A.1.2 Physics

• Physics agents support difficult concepts and simulations; work includes NEWTON’s physical-reasoning pipeline and benchmark and a case study of passing an introductory calculus-based physics course.Wang et al. (2023) introduce NEWTON; Kortemeyer (2023) describes the course case study.Source: §A.1.2 · p. 23
• Physics Reasoner uses a comprehensive formula set and detailed checklists in a three-stage process of problem analysis, formula retrieval, and guided reasoning.Pang et al. (2024) present it as a knowledge-augmented framework for physics problem solving.Source: §A.1.2 · p. 23
• Ma et al. (2024a) describe SGA, a bilevel optimization framework for physical scientific discovery that generates and revises hypotheses using an exploit-and-explore strategy.Source: §A.1.2 · p. 23

A.1.3 Chemistry

• Chemistry agents support explanations and synthesis: ChemCrow plans and executes syntheses (M. Bran et al., 2024), ChemAgent emphasizes reasoning and grounding (Yu et al., 2024), and Curie adds rigor modules (Kon et al., 2025).Curie’s modules target reliability, systematic control, and interoperability; other work explores complex chemical discovery and agent-based tool use with flexible planning (Yang et al., 2024d; Sprueill et al., 2024).Source: §A.1.3 · pp. 23, 25

A.1.4 Biology

• Biology agents enable protein analysis, modeling, and engineering through systems including ProtChat, ProtAgents (Ghafarollahi and Buehler, 2024a), and TourSynbio-Agent.ProtAgents assigns User, Planner, Assistant, Critic, and Group Chat Manager roles; TourSynbio-Agent supports mutation analysis, inverse folding, and visualization.Source: §A.1.4 · p. 25

A.1.5 General Scientific Discovery

• Scientific-discovery agents help analyze datasets and use domain tools for research tasks, while Agent Laboratory automates literature review, experimentation, and report writing (Lála et al., 2023; Chen et al., 2025a; Schmidgall et al., 2025).Aviary covers DNA construct manipulation, literature-based research questions, and protein stability engineering (Narayanan et al., 2024); SciAgent and SciToolAgent augment scientific reasoning with domain-specific tools (Ma et al., 2024c; Chen et al., 2025a).Source: §A.1.5 · p. 25

A.2.1 Reading

• Reading agents adapt and simplify texts for comprehension and serve as academic reading companions.ExpertEase adapts documents to grade-specific audiences, while AgentSimp uses multiple role-specific agents to support coherence and accessibility; reading companions are reported to improve engagement and understanding of complex qualitative texts (Chen and Leitch, 2024).Source: §A.2.1 · p. 25

A.2.2 Writing

• Writing support spans grammatical-error correction and essay scoring, task-tuned Weaver, interactive CAELF feedback, and persona-based Debate-to-Write planning (Ye et al., 2024b, 2025a; Su et al., 2025; Wang et al., 2024c; Hong et al., 2024; Hu et al., 2025b).Weaver outperforms generalist LLMs such as GPT-4 on human-like narrative generation and supports retrieval-augmented generation and function calling (Wang et al., 2024c). CAELF combines teaching-assistant evaluations with teacher arbitration so students can contest grades; Debate-to-Write agents collaboratively discuss ideas and plan arguments (Hong et al., 2024; Hu et al., 2025b).Source: §A.2.2 · pp. 25–26

A.2.3 Translation

• Translation agents combine specialized tools and collaboration: Agent-SiMT dynamically balances reading and generation actions, while TransAgents models human translation teams’ collaborative reasoning (Guo et al., 2024a; Wu et al., 2024a).The passage describes Agent-SiMT as achieving state-of-the-art simultaneous translation performance and TransAgents as addressing cultural nuance and stylistic challenges.Source: §A.2.3 · p. 26

A.2.4 Storytelling

• Storytelling agents scaffold children’s narrative creation with embodied multimodal interaction, prompt-to-narrative generation, and long-sequence story illustration (Li and Xu, 2023; Sohn et al., 2024).STARie uses speech synthesis and facial animation; StoryAgent combines top-down drafting with bottom-up asset generation; LLaMS uses a Story-Adapter for long image-sequence illustration.Source: §A.2.4 · p. 26

A.2.5 Speaking

• Speaking agents support realistic language practice through contextual feedback, task-oriented spoken dialogue, embodied verbal and nonverbal cues, and multimodal simulations (Pan et al., 2024; Zhang et al., 2024a).Multi-Agent Tuning supports simulations with up to 25 agents, including drama creation and audio-novel generation (Liang et al., 2024).Source: §A.2.5 · p. 26

A.3 Agent for Professional Development

• Professional-development agents aim to deliver scalable, adaptive, context-aware learning across medical, computer-science, and law training.The section surveys domain-specific approaches to professional training.Source: §A.3 · p. 27

A.3.1 Medical Education

• Agent Hospital agents improved medical reasoning and MedQA performance after treating thousands of simulated patients (Li et al., 2024b; Jin et al., 2021).The simulation environment lets LLM-driven agents evolve through autonomous interactions.Source: §A.3.1 · p. 27
• Medical-education LLMs can support personalized curricula, adaptive learning plans, and dynamic assessments, while requiring attention to bias, misinformation, and privacy.Source: §A.3.1 · p. 27
• MEDCO simulates real-world medical training through collaboration among virtual patients, expert physicians, and radiologists, supporting interdisciplinary learning and peer interaction (Wei et al., 2024).Source: §A.3.1 · p. 27
• openCHA integrates external resources and orchestrates multi-step problem-solving for complex healthcare queries, emphasizing tool use and action planning (Abbasian et al., 2023; Ye et al., 2024a).Source: §A.3.1 · p. 27
• Fantasy narrative games for adolescent health education can provide engaging, doctor-validated content that enhances knowledge retention through gamification.Source: §A.3.1 · p. 27

A.3.2 Computer Science Education

• CS agents offer personalized guidance on coding, debugging, and CS principles, while Iris provides context-aware tutoring within Artemis (Bassner et al., 2024).The section also describes agent-supported coding efficiency and Learning-by-Teaching as a pedagogical strategy.Source: §A.3.2 · p. 27
• CodeAgent (Zhang et al., 2024b) supports repo-level generation with WebSearch and DocSearch; ToolCoder, SWE-agent, AgentCoder, and MapCoder are reported to enhance coding efficiency.The latter systems are associated with Jin et al. (2024a), Wang et al. (2025a), and Frankford et al. (2024).Source: §A.3.2 · p. 27
• Learning-by-Teaching uses LLM agents’ contextual conversation and learning from demonstrations as a pedagogical strategy for computer science education (Chen et al., 2024a).Source: §A.3.2 · p. 27

A.3.3 Law Education

• LLM agents can support law education through judicial interpretation, moot court simulation, and case analysis (Chen et al., 2024b; Nelson, 2024; Lai et al., 2024; Yuan et al., 2024).Source: §A.3.3 · p. 27
• LawBench and LegalBench evaluations find difficulties in legal knowledge application and judicial aid, while LegalAgentBench identifies weaknesses in multi-hop reasoning and defense statement writing (Fei et al., 2023; Guha et al., 2023; Li et al., 2024a).The section concludes that agents need significant improvement to assist effectively with complex legal tasks.Source: §A.3.3 · p. 27
• DeliLaw combines legal and case retrieval for research, case analysis, statutory interpretation, and mock consultations; LawLuo uses multi-agent retrieval-augmented generation for multi-turn consultations (Sun et al., 2024).LawLuo is described as improving personalization and ambiguity handling.Source: §A.3.3 · pp. 27–28
• AgentCourt and AgentsCourt simulate courtroom interactions and judicial decision-making, while AgentsBench extends legal education with multi-agent reasoning and case analysis (Chen et al., 2024b; He et al., 2024b; Jiang and Yang, 2024).Source: §A.3.3 · p. 28

B Datasets & Benchmarks

• ASSIST09 (Feng et al., 2009) and Junyi support K-12 math knowledge tracing; EduAgent (Xu et al., 2024a) enables adaptive learning.Virtual Teacher (Xu et al., 2024b) and MathCCS (Zhang et al., 2025d) assess error correction and detection in math learning.Source: B Datasets & Benchmarks · p. 28
• FABRIC (Han et al., 2023), EssayJudge (Su et al., 2025), and EXCGEC (Ye et al., 2024b) focus on essay feedback and generation; MultiSim and Wang et al. (2024b) benchmark multilingual translation and storytelling.The latter benchmarks extend educational LLM evaluation beyond English-language education.Source: B Datasets & Benchmarks · p. 28
• ScienceAgentBench (Chen et al., 2024d) and TheoremExplainBench (Ku et al., 2025) assess scientific reasoning and theorem explanation, while ML-Bench (Tang et al., 2023) and MLAgentBench (Huang et al., 2023c) target machine-learning education.These domain-specific benchmarks assess specialized knowledge, problem-solving, and interactive learning for subject instruction.Source: B Datasets & Benchmarks · p. 28
• LawBench (Fei et al., 2023), LegalBench (Guha et al., 2023), and AgentCourt (Chen et al., 2024b) evaluate legal applications; MedBench (Cai et al., 2024) and OmniMedVQA (Hu et al., 2024) test clinical reasoning and medical retrieval.For computer science, SWE-Bench and Programming Feedback assess code generation, debugging, and software-engineering instruction.Source: B Datasets & Benchmarks · p. 28
• Table 2 organizes publicly available educational-agent datasets and benchmarks by goal, target users, subject, education level, language, modality, and size.The authors present the collection as a resource intended to support research on LLM agents for education.Source: B Datasets & Benchmarks · p. 28, Table 2 · p. 29