Pith. sign in

REVIEW 4 major objections 5 minor 25 references

AI-Powered Math Tutoring: Platform for Personalized and Adaptive Education

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A 'Tutor Prompt' using Socratic questioning makes LLM math tutors reach more correct solutions while revealing the answer far less often in simulated dialogues, validating the guided-tutoring design of a multi-agent platform.

desk verdict A reasonable multi-agent tutor architecture paper whose central platform claims outrun an evaluation that only tests one prompt variant on a simulated dialogue benchmark. read the letter →

arxiv 2507.12484 v1 pith:CRBSD2X2 submitted 2025-07-14 cs.AI cs.MA

Jaros{l}aw A. Chudziak , Adam Kostka This is my paper
classification cs.AIcs.MA
keywords ArtificialIntelligenceMulti-AgentSystemsLargeLanguageModelsTutoringPersonalizedLearningMathEducationGraphRAGSocratic
verification ladder T0 review T1 audit T2 compute T3 formal

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 is trying to establish that an AI math tutor does not have to be a reactive answer machine: it can be built as a multi-agent system that guides a student through Socratic questioning, personalises the session through dual memory, pulls in textbook content through graph-based retrieval, and plans courses as a directed acyclic graph. The central, benchmark-validated claim is that a carefully engineered 'Tutor Prompt' makes large language models tutor more effectively in simulated dialogues than a basic prompt, raising the rate of successful solutions while sharply lowering the rate at which the model simply tells the student the answer. The authors present this as evidence that their platform's core design holds promise for structured, individualized, tool-assisted mathematics learning, and they identify the absence of a real-student trial as the main remaining limitation. A sympathetic reader would care because mathematics education needs tutors that support deep understanding, and a reusable prompting and architecture strategy would be a concrete step in that direction.

What carries the argument

The load-bearing mechanism is the Tutor Prompt, an instruction set that pushes the LLM toward Socratic dialogue: asking pointed questions, offering scaffolds, and holding back the final answer. It is evaluated by two metrics defined on the MathDial dataset: Success@K counts how often the simulated student reaches a correct solution within K dialogue turns, and Telling@K counts how often the tutor's own turns give the answer away; a good tutor should raise the first and lower the second. Around this prompt, the platform is organised as a multi-agent system with a GPT-4o Tutor Agent orchestrating tools via a ReAct-style loop, a memory dispatcher separating long-term student traits from working-memory session context, graph-based textbook retrieval, a symbolic solver, and a directed-acyclic-graph course planner. This combination is what the authors claim turns reactive answer-giving into structured, personalised guidance.

What would settle it

A randomised classroom study comparing the Tutor Prompt against the Base Prompt on the same topic, with the same LLM and equivalent time on task, would settle the central claim: if students tutored by the Tutor Prompt show no better post-test scores than students tutored by the Base Prompt, then the benchmark's promise of guided tutoring does not translate into learning outcomes. A cheaper failure test is to reproduce the MathDial evaluation with the released prompts and a held-out dialogue split; if the Tutor Prompt no longer beats the Base Prompt on Success@K and Telling@K, the reported result does not replicate.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the way an LLM tutor is prompted changes its pedagogical behaviour in a measurable way. Using MathDial, a dataset of simulated tutor-student math dialogues, the authors compared their pedagogically informed 'Tutor Prompt'—which emphasises Socratic questioning, scaffolding, and delaying the answer—against MathDial's 'Base Prompt' across GPT-4o and GPT-4o-mini. For both models the Tutor Prompt achieved higher Success@K (dialogues reaching a correct solution within K turns) and far lower Telling@K (turns that directly reveal the answer). They also benchmarked several LLMs with access to a symbolic solver on the MathDial problem set: o3-mini(high) and Claude 3.5 Sonnet reached 90.00% accuracy, Gemini 2.0 Flash 88.67%, and GPT-4o and GPT-4o-mini about 78.67% and 77.33%, leading them to select o3-mini(high) for the task-creation component. The paper presents these results as validating the guided-tutoring prompt strategy and informing the model choices inside the broader multi-agent platform.

Load-bearing premise

The load-bearing premise is that Success@K and Telling@K measured on MathDial's simulated dialogues are a valid proxy for real tutoring quality and real learning gains; the paper's own Section 6 concedes that no evaluation with real students was performed.

Editorial extensions

If this is right

  • If the Tutor Prompt result holds, the same prompt strategy can be dropped into other LLM-based tutors to reduce answer-giving and increase guided problem-solving in simulated settings.
  • The model benchmark gives a concrete model choice—o3-mini(high) with a symbolic solver at 90.00% on MathDial problems—for anyone building a math exercise generator.
  • The open-source multi-agent architecture demonstrates that memory, graph retrieval, course planning, and tutoring can be combined into one modular platform rather than as separate tools.
  • The system's dependence on the underlying LLM means future improvements in model reasoning should directly improve both tutoring guidance and task generation.
  • Because the evaluation is simulation-only, the platform's real-world effect on learning remains untested until a student trial is run.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension the authors leave implicit: Telling@K and Success@K could be used as cheap, repeatable screening metrics for tutoring prompts before any classroom study, so the same benchmark can drive iterative prompt improvement.
  • The graph-based retrieval and Socratic-prompting components are justified by separate arguments, but the paper does not test the two together; combining them to ground Socratic hints in textbook material is a natural experiment to run.
  • The model accuracy ranking comes from a single dataset with access to a symbolic solver, so whether o3-mini(high) remains the best choice for harder, multi-step mathematical reasoning is not established by the paper.
  • If low Telling@K truly predicts better learning, the paper's metric could become a general proxy for pedagogical quality in AI-tutor research, but that link still needs real-student validation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces a multi-agent AI tutoring platform for mathematics that combines a Tutor Agent (GPT-4o) using Socratic prompting, dual-memory personalization (long-term and working memory), GraphRAG-based textbook retrieval, DAG-structured course planning, and a task generation module. The authors evaluate two aspects: (i) the Tutor Agent's guiding behavior on the MathDial benchmark, comparing a custom 'Tutor Prompt' against MathDial's 'Base Prompt' using Success@N and Telling@N metrics, and (ii) the problem-solving accuracy of several LLMs on the MathDial problem set, which informs the choice of o3-mini(high) for the Task Creation component. The central claim is that the Tutor Prompt enables more effective guided tutoring (higher Success@N, lower Telling@N) than the base prompt.

Significance. If the reported results are statistically reliable, the paper offers a modest but useful empirical result: a carefully engineered Socratic prompt can steer an LLM tutor toward scaffolding rather than answer-giving in simulated dialogues. The open-source release and the use of an external benchmark (MathDial) are strengths, and the model-selection procedure for the task-creation component is a reasonable engineering decision. However, the paper's broader claims about a validated tutoring platform that enables students to 'learn new topics,' 'revise for exams effectively,' and receive 'adaptive and personalized feedback' extend far beyond what the evaluation actually tests: the benchmark evaluation isolates only the Tutor Agent's prompting, not the platform's distinctive components (memory, GraphRAG, DAG planning, task creation), and no real-student study is included. The paper's significance therefore rests on a narrow prompt-engineering result plus an unvalidated architectural proposal.

major comments (4)
  1. [Section 5] The two central metrics, Success@N and Telling@N, are never defined. The reader cannot assess what 'success' or 'telling' means operationally, how the interaction length K is varied, or whether these metrics are the ones used in the MathDial benchmark or newly introduced. The claim that the Tutor Prompt 'significantly outperformed' the Base Prompt requires definition of the metrics, sample sizes, error bars or significance tests, and details of the evaluation protocol (e.g., number of dialogues, temperature settings, output parsing). Without these, the headline result of the paper is not reproducible.
  2. [Section 5 and Section 6] The evaluation exercises only the Tutor Agent's prompting strategy in a simulated, tool-free MathDial dialogue; it does not test the platform's claimed differentiators: dual-memory personalization, GraphRAG retrieval, DAG-based course planning, or the Task Creation module. Yet the abstract and conclusion attribute system-level learning benefits ('learn new topics,' 'revise for exams effectively,' 'adaptive and personalized feedback'). Section 6 explicitly concedes the 'lack of evaluation with real students in learning environments' and that the 'pedagogical effectiveness of courses generated through GraphRAG for students is yet to be evaluated empirically.' This concession directly undercuts the platform-level claims; the conclusions should be restricted to what the benchmark can support, or a real-student pilot should be added.
  3. [Section 5] The model-selection result for the Task Creation component is obtained on the MathDial problem set, which is also the benchmark used for the Tutor Agent evaluation. If the same problem instances are used both to select a model (o3-mini(high)) and to validate the tutoring approach, there is a risk of optimistic bias or unintended selection on the test set. The paper should clarify whether the problem sets are disjoint, and if not, report the selection process as exploratory and temper any claims of benchmark-validated component choice.
  4. [Section 5] The accuracy comparison for model selection (o3-mini(high) and Claude 3.5 Sonnet at 90.00%, Gemini 2.0 Flash at 88.67%, GPT-4o at 78.67%, GPT-4o-mini at 77.33%) is reported as point estimates with no error bars, confidence intervals, or significance tests, despite the small gaps (e.g., 90.00% vs. 88.67%). The selection of o3-mini(high) over Claude 3.5 Sonnet is also not justified: if both have equal accuracy, the stated criterion ('top performance and step-decomposition ability') should be supported by explicit evidence about step-decomposition quality, which is not presented.
minor comments (5)
  1. [Section 4.1] The statement that GPT-4o was chosen for the Tutor Agent 'for its conversational capabilities, proficiency in tool use, and low latency' is not backed by any comparison or citation; a brief justification or reference would improve the presentation.
  2. [Section 5] Figure 4 is described only in the caption and a paragraph; the axes, the definition of K, and the difference between the plotted curves are not explained in the text. Please add a full figure description in the caption or body.
  3. [References] Reference [2] is listed as 'proceedings forthcoming' and reference [10] as a PACLIC paper; please update these with publication details if they are now available, or note that they are preprints.
  4. [Section 6] The phrase 'noted by reviewers' in 'a significant limitation noted by reviewers' is unusual for a paper not under review; rephrase to a plain statement of the limitation.
  5. [Throughout] The paper would benefit from consistency in the use of 'Tutor Agent' vs 'Tutoring Agent' (both appear) and in the capitalization of 'Task Creation' vs 'task creation'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the prompting evaluation is anchored to the external MathDial benchmark and the broader learning claims are explicitly conceded as unevaluated in Section 6, which is a validity gap rather than a circular derivation.

full rationale

The paper's derivation chain is not circular. The central evaluation in Section 5 compares a pedagogically designed Tutor Prompt against MathDial's Base Prompt on the external MathDial dataset; the prompt is not fitted to the Success@N or Telling@N metrics, and no equation or parameter is defined in terms of the outcome it is said to validate. The model-selection step for Task Creation uses the same MathDial problem set to choose o3-mini(high), but this is a disclosed design choice rather than a prediction claim, and it does not reduce to a fitted input being reported as an independent result. The two self-citations (references [2] and [10]) appear only as peripheral related-work support and are not load-bearing for any core claim. Section 6 explicitly concedes that real-student learning gains and GraphRAG course effectiveness are unevaluated; that is an external-validity limitation, not circularity. No step in the paper's argument is equivalent to its inputs by construction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The ledger is light: the paper builds from existing building blocks (multi-agent orchestration, GraphRAG, LLMs) rather than postulating new scientific entities. The main unexamined inputs are the pedagogical assumptions and the benchmark-proxy assumption, plus hand-chosen prompt and model assignments. No numbers are fitted to data in a derivation sense.

free parameters (3)
  • Tutor Prompt design
    Hand-crafted Socratic prompting template; the central evaluation claim (guidance effectiveness) is defined relative to this prompt, which is a design choice rather than a derived quantity.
  • Component model assignments = GPT-4o (tutor); o3-mini(high) (task creation)
    Model choices for tutor and task-creation agents were selected based on the Section 5 MathDial evaluation, making the measured platform performance contingent on these hand-chosen assignments.
  • GraphRAG index construction parameters
    The paper does not specify indexing, graph extraction, or retrieval parameters for GraphRAG, which are hand-chosen in practice and affect the retrieved context that grounds tutor responses.
assumptions (4)
  • domain assumption MathDial benchmark dialogues are a valid proxy for real tutor-student interactions
    Section 5 evaluates the tutoring prompt on simulated dialogues and generalizes to tutoring effectiveness without real-student data; Section 6 admits real-student evaluation is still missing.
  • domain assumption Socratic questioning and scaffolding improve learning outcomes
    The Tutor Prompt's pedagogical premise relies on established learning-science claims (refs 15, 19), imported as background rather than validated by this paper's experiments.
  • domain assumption GraphRAG's graph structure better represents educational content than vector RAG for tutoring
    Stated in Section 2 citing ref 7; the paper presents no own comparison of GraphRAG against plain RAG in its evaluation.
  • domain assumption LLM outputs are reliable enough for production tutoring with memory and tool components
    The system is built on GPT-4o and o3-mini outputs without guardrails or reliability analysis; Section 6 acknowledges dependence on underlying LLM capabilities and biases.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AI-Powered Math Tutoring: Platform for Personalized and Adaptive Education." pith.science (2026). https://pith.science/paper/CRBSD2X2

@misc{pith2026250712484,
  author       = {Pith},
  title        = {Pith review of: AI-Powered Math Tutoring: Platform for Personalized and Adaptive Education},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CRBSD2X2}},
  note         = {Machine review of arXiv:2507.12484}
}
read the original abstract

The growing ubiquity of artificial intelligence (AI), in particular large language models (LLMs), has profoundly altered the way in which learners gain knowledge and interact with learning material, with many claiming that AI positively influences their learning achievements. Despite this advancement, current AI tutoring systems face limitations associated with their reactive nature, often providing direct answers without encouraging deep reflection or incorporating structured pedagogical tools and strategies. This limitation is most apparent in the field of mathematics, in which AI tutoring systems remain underdeveloped. This research addresses the question: How can AI tutoring systems move beyond providing reactive assistance to enable structured, individualized, and tool-assisted learning experiences? We introduce a novel multi-agent AI tutoring platform that combines adaptive and personalized feedback, structured course generation, and textbook knowledge retrieval to enable modular, tool-assisted learning processes. This system allows students to learn new topics while identifying and targeting their weaknesses, revise for exams effectively, and practice on an unlimited number of personalized exercises. This article contributes to the field of artificial intelligence in education by introducing a novel platform that brings together pedagogical agents and AI-driven components, augmenting the field with modular and effective systems for teaching mathematics.

Figures

Figures reproduced from arXiv: 2507.12484 by the authors.

Figure 1
Figure 1. Guided tutoring interaction showing Socratic questioning and personalized sup￾port components. However, many current LLM tutors provide direct answers, often overlook￾ing the crucial cognitive steps necessary for effective learning and hindering deep understanding [16], lack robust integration with course materials [12], and offer limited dynamic adaptation [18]. A gap exists for AI tutors that foster deep learning … view at source ↗
Figure 2
Figure 2. Course creation using a DAG structure for structured learning and targeted practice. tutoring approach. Previous research on ITS architectures [19] and RAG assis￾tants [20] has contributed to our work. Retrieval-Augmented Generation (RAG) is crucial for knowledge retrieval, grounding LLMs knowledge [12]. We use GraphRAG [4] based on the argument that its graph structure better represents educational content relation… view at source ↗
Figure 3
Figure 3. System architecture: Information flow between Student, Tutor Agent, Mem￾ory, Specialist Agents (Research, Planning, etc.), and Tools (Retrieval, Task Creation, Solver, etc.). proactively offer targeted hints, scaffold the problem differently, or provide cor￾rective feedback when the error occurs. On the same principle, if LTM notes that a student prefers visual explanations, aids such as function plotting will be us… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Guided Tutoring Benchmark (Mathdial): Tutor Prompt yields higher Suc￾cess@K and lower Telling@K than Base Prompt, validating its guidance effectiveness. tent, student-specific data (e.g., prior knowledge, misconceptions, learning pref￾erences, goals), while working mem…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 10 canonical work pages

  1. [1]

    In: Proc

    Chen, Y., Ding, N., Zheng, H.T., Liu, Z., Sun, M., Zhou, B.: Empow- ering private tutoring by chaining large language models. In: Proc. 33rd ACM Int. Conf. Inf. Knowl. Manag. (CIKM ’24). pp. 354–364. ACM (2024). https://doi.org/10.1145/3627673.3679665

    arXiv 2024

  2. [2]

    In: Proc

    Chudziak, J.A., Wawer, M.: Elliottagents: A natural language-driven multi-agent system for stock market analysis and prediction. In: Proc. 38th Pacific Asia Conf. Lang. Inf. Comput. (PACLIC 38). Tokyo, Japan (2024), presented at PACLIC 38; proceedings forthcoming

    work page 2024

  3. [3]

    In: Proc

    Chudziak,J.A.,Cinkusz,K.:TowardsLLM-augmentedmultiagentsystemsforagile software engineering. In: Proc. 39th IEEE/ACM Int. Conf. Autom. Softw. Eng. (ASE ’24). pp. 2476–2477 (2024)

    work page 2024

  4. [4]

    Edge, D., Trinh, H., Cheng, N., Bradley, J., Chao, A., Mody, A., Truitt, S., Metropolitansky, D., Ness, R.O., Larson, J.: From local to global: A graph rag ap- proach to query-focused summarization (2025), https://arxiv.org/abs/2404.16130

    arXiv 2025

  5. [5]

    Graesser, A.C., Wiemer-Hastings, K., Wiemer-Hastings, P., Kreuz, R.: Auto- Tutor: A simulation of a human tutor. Cogn. Syst. Res. 1(1), 35–51 (1999). https://doi.org/10.1016/S1389-0417(99)00005-4

    work page doi:10.1016/s1389-0417(99)00005-4 1999

  6. [6]

    Guan, X., Zhang, L.L., Liu, Y., Shang, N., Sun, Y., Zhu, Y., Yang, F., Yang, M.: rstar-math: Small llms can master math reasoning with self-evolved deep thinking (2025), https://arxiv.org/abs/2501.04519

    arXiv 2025

  7. [7]

    graphrag: A systematic evaluation and key insights (2025), https://arxiv.org/abs/2502.11371

    Han, H., Shomer, H., Wang, Y., Lei, Y., Guo, K., Hua, Z., Long, B., Liu, H., Tang, J.: Rag vs. graphrag: A systematic evaluation and key insights (2025), https://arxiv.org/abs/2502.11371

    arXiv 2025

  8. [8]

    Heffernan, N., Heffernan, C.: The ASSISTments ecosystem: Building a plat- form that brings scientists and teachers together for minimally invasive research on human learning and teaching. Int. J. Artif. Intell. Educ. 24 (Dec 2014). https://doi.org/10.1007/s40593-014-0024-x

    doi:10.1007/s40593-014-0024-x 2014

Show all 25 references
  1. [9]

    Research in Mathematics Ed- ucation 0(0), 1–22 (2023)

    van Hoeve, M., Doorman, M., Veldhuis, M.: Fostering a growth mindset in sec- ondary mathematics classrooms in the netherlands. Research in Mathematics Ed- ucation 0(0), 1–22 (2023). https://doi.org/10.1080/14794802.2023.2241433

    2023

  2. [10]

    In: Proc

    Kostka, A., Chudziak, J.A.: Synergizing logical reasoning, long-term memory, and collaborative intelligence in multi-agent llm systems. In: Proc. 38th Pacific Asia Conf. Lang. Inf. Comput. (PACLIC 38). Tokyo, Japan (2024)

    2024

  3. [11]

    Artificial Intelligence Review 57 (08 2024)

    Kumar, P.: Large language models (llms): survey, technical frame- works, and future challenges. Artificial Intelligence Review 57 (08 2024). https://doi.org/10.1007/s10462-024-10888-y

    2024 doi

  4. [12]

    In: Artificial Intelligence in Education

    Li, X., Henriksson, A., Duneld, M., Nouri, J., Wu, Y.: Supporting teaching- to-the-curriculum by linking diagnostic tests to curriculum goals: Using text- book content as context for retrieval-augmented generation with large lan- guage models. In: Artificial Intelligence in Ed...

    2024 doi

  5. [13]

    Liang, L., Sun, M., Gui, Z., Zhu, Z., Jiang, Z., Zhong, L., Qu, Y., Zhao, P., Bo, Z., Yang, J., Xiong, H., Yuan, L., Xu, J., Wang, Z., Zhang, Z., Zhang, W., Chen, 8 J. A. Chudziak, A. Kostka H., Chen, W., Zhou, J.: Kag: Boosting llms in professional domains via knowledge augme...

    2024 arXiv

  6. [14]

    2024 IEEE Conference on Artificial Intelligence (CAI) pp

    Liu, Z., Yin, S.X., Lee, C., Chen, N.F.: Scaffolding language learn- ing via multi-modal tutoring systems with pedagogical instructions. 2024 IEEE Conference on Artificial Intelligence (CAI) pp. 1258–1265 (2024), https://api.semanticscholar.org/CorpusID:268889625

    2024

  7. [15]

    In: Bouamor, H., Pino, J., Bali, K

    Macina, J., Daheim, N., Chowdhury, S., Sinha, T., Kapur, M., Gurevych, I., Sachan, M.: MathDial: A dialogue tutoring dataset with rich pedagogical prop- erties grounded in math reasoning problems. In: Bouamor, H., Pino, J., Bali, K. (eds.) Findings of the Assoc. Comput. Lingui...

    2023 doi

  8. [16]

    In: Vlachos, A., Augenstein, I

    Macina, J., Daheim, N., Wang, L., Sinha, T., Kapur, M., Gurevych, I., Sachan, M.: Opportunities and challenges in neural dialog tutoring. In: Vlachos, A., Augenstein, I. (eds.) Proc. 17th Conf. Eur. Chapter Assoc. Comput. Lin- guist. (EACL 2023). pp. 2357–2372. ACL, Dubrovnik,...

    2023 doi

  9. [17]

    Scientific Reports 14(1) (Dec 2024)

    Mishra, T., Sutanto, E., Rossanti, R., Pant, N., Ashraf, A., Raut, A., Uwabareze, G., Oluwatomiwa, A., Zeeshan, B.: Use of large language models as artificial intel- ligence tools in academic research and publishing among global clinical researchers. Scientific Reports 14(1) (...

    2024 doi

  10. [18]

    Park, M., Kim, S., Lee, S., Kwon, S., Kim, K.: Empowering personalized learning through a conversation-based tutoring system with student modeling. In: Ext. Abstracts CHI Conf. Hum. Factors Comput. Syst. (CHI EA ’24). pp. 1–10. ACM (May 2024). https://doi.org/10.1145/3613905.3651122

  11. [19]

    In: Bouamor, H., Pino, J., Bali, K

    Sonkar, S., Liu, N., Mallick, D., Baraniuk, R.: CLASS: A design frame- work for building intelligent tutoring systems based on learning science princi- ples. In: Bouamor, H., Pino, J., Bali, K. (eds.) Findings of the Assoc. Com- put. Linguist.: EMNLP 2023. pp. 1941–1961. ACL, ...

    2023 doi

  12. [20]

    Taneja, K., Maiti, P., Kakar, S., Guruprasad, P., Rao, S., Goel, A.K.: Jill watson: A virtual teaching assistant powered by chatgpt (2024), https://arxiv.org/abs/2405.11070

    2024 arXiv

  13. [21]

    Nature 625(7995), 476–482 (Jan 2024)

    Trinh, T.H., Wu, Y., Le, Q.V., He, H., Luong, T.: Solving olympiad geometry without human demonstrations. Nature 625(7995), 476–482 (Jan 2024). https://doi.org/10.1038/s41586-023-06747-5, https://www.nature.com/articles/s41586-023-06747-5#citeas

    2024 doi

  14. [22]

    Viswanathan, N., Meacham, S., Adedoyin, F.F.: Enhancement of online education system by using a multi-agent approach. Comput. Educ. Artif. Intell.3, 100057 (2022). https://doi.org/10.1016/j.caeai.2022.100057

    2022

  15. [23]

    Wang, S., Xu, T., Li, H., Zhang, C., Liang, J., Tang, J., Yu, P.S., Wen, Q.: Large language models for education: A survey and outlook (2024), https://arxiv.org/abs/2403.18105

    2024 arXiv

  16. [24]

    Wang, X., Xu, X., Zhang, Y., Hao, S., Jie, W.: Exploring the impact of artificial intelligence application in personalized learning environments: thematic analysis of undergraduates’ perceptions in China. Humanit. Soc. Sci. Commun.11(1) (Dec 2024). https://doi.org/10.1057/s415...

    2024 doi

  17. [25]

    Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., Cao, Y.: React: Synergizing reasoning and acting in language models (2023), https://arxiv.org/abs/2210.03629

    2023 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.