Pith. sign in

REVIEW 4 major objections 8 minor 1 cited by

GraphRAG-Induced Dual Knowledge Structure Graphs for Personalized Learning Path Recommendation

T0 review · 4 major / 8 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 Similarity-based fallback improves learning path recommendations

desk verdict Dual prerequisite/similarity graphs plus a similarity-fallback RL agent is a genuinely new combination, but the SOTA claim rests on an in-loop DKT evaluator that also serves as the reward, so the evidence is conditional until they provide held-out validation. read the letter →

arxiv 2506.22303 v2 pith:BMHNU4VC submitted 2025-06-27 cs.IR

Xinghe Cheng , Zihan Zhang , Jiapu Wang , Liangda Fang , Chaobo He , Quanlong Guan , Shirui Pan , Weiqi Luo This is my paper
classification cs.IR
keywords learningpathrecommendationknowledgeconceptgraphprerequisiterelationshipssimilarityreinforcementRAGtracinglargelanguagemodels
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 tries to show that learning path recommendation improves when the learner's route is guided by two kinds of relations between knowledge concepts: prerequisite order and similarity. When a learner gets stuck on a prerequisite concept, a similarity agent inserts closely related concepts to help them discriminate and move on. The authors build KnowLP, which generates both graphs from concept names using an LLM pipeline, then trains two cooperating reinforcement-learning agents plus a difficulty matcher. On three educational datasets, KnowLP reports higher learning-gain scores than previously published methods, with the largest gap on longer paths. If this holds, expert-annotated prerequisite graphs would no longer be required for good path recommendation, and blocked progress would be less disruptive.

What carries the argument

The load-bearing object is the dual knowledge structure graph, a pair of graphs over knowledge concepts: one with prerequisite edges and one with similarity edges. The paper's EDU-GraphRAG builds both from the names of the concepts alone, so no expert annotation is needed; TextGrad is used to iteratively refine the LLM-generated explanations that ground the graph extraction. The other central mechanism is the DLRL multi-agent policy: a P-Agent picks the next concept by prerequisite relations, an S-Agent supplies a fallback sub-path of confusable similar concepts when the learner's predicted mastery gain falls below a threshold, and a D-Agent matches exercises to the learner's current mastery level. The environment is DIMKT, a difficulty-aware knowledge tracing model, and the agents are trained with PPO toward the reward of predicted mastery improvement.

What would settle it

Run a controlled trial in which learners are given either KnowLP-generated paths or a baseline path and then take a common post-test on the target concepts; if the post-test gains do not reproduce the ordering predicted by DKT mastery, the central claim fails. More directly, compare KnowLP's predicted improvement with actual post-test improvement on the same learners and check whether the similarity-agent interventions correspond to real recovery.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that dual knowledge structure graphs—prerequisite and similarity—are enough to make learning path recommendation both stronger and more adaptive. The central mechanism is an agent-switching rule: a prerequisite agent plans the next knowledge concept, and when the learner's predicted mastery of the last concept barely improves, a similarity agent takes over and weaves in concepts that are easily confused with the target. The similarity edges are generated by EDU-GraphRAG, which uses a language model to write refined explanations of each knowledge concept and then extracts a relation graph from them, reducing hallucinations through TextGrad feedback. The paper reports that KnowLP outperforms all compared baselines on Junyi, MOOCCubeX, and ASSISTments2009, measured by the predicted fraction of learning goals achieved, and the ablation shows the similarity agent is responsible for much of the gain.

Load-bearing premise

The evaluation and the reinforcement learning reward both treat a knowledge tracing model's predicted mastery as the true measure of learning, so if those predictions are optimistic or biased, the reported gains could be artifacts of optimizing that predictor rather than real learning improvement.

Editorial extensions

If this is right

  • Learning path recommenders can run on datasets with no prerequisite graph at all, as long as knowledge concept names are available.
  • Longer learning paths, which lost quality in most baselines, become the regime where KnowLP shows its largest gains, indicating the similarity fallback dampens cascading blockage.
  • The generated dual graphs could be reused separately, for example as a similarity graph for designing interleaved practice or as a prerequisite graph for other educational tools.
  • The switching threshold for activating the similarity agent provides a single interpretable control for balancing strict prerequisite order against recovery flexibility.

Reading between the lines

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

  • The paper does not test the method with real students; a post-test study would settle whether DKT-predicted gains translate to actual learning.
  • Beyond the paper's RL recommender, the generated similarity graph could stand alone as a tool for designing interleaved practice schedules, a strategy known to help learners distinguish similar concepts.
  • The fixed switching threshold suggests a personalized extension where the threshold adapts to each learner's historical stuck rate, which the paper leaves unexplored.
  • The generated graph artifacts could also serve downstream educational tasks, such as automatic prerequisite prediction or concept tagging, independent of the path recommendation module.
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 / 8 minor

Summary. The paper proposes KnowLP, a learning path recommendation framework that combines an LLM-based graph generation module (EDU-GraphRAG) with a multi-agent reinforcement learning module (DLRL). EDU-GraphRAG uses TextGrad-refined LLM explanations to construct prerequisite and similarity graphs among knowledge concepts from dataset-specific KC names. DLRL comprises a prerequisite agent (PPO-based) and a similarity agent that activates when progress is low, together with a difficulty agent for exercise selection, all interacting with a DIMKT knowledge-tracing environment. The authors evaluate on Junyi, MOOCCubeX, and ASSISTments2009 using Ep, the improvement in DKT-predicted mastery over a recommended path, and report that KnowLP outperforms eight baselines at all path lengths and yields larger gains at step=20. An ablation shows performance drops without the similarity agent, and a simulation study on a DKT-based simulator reports further improvements over three strong baselines.

Significance. If the reported results are validated, the paper makes a useful contribution: it addresses the cost and brittleness of prerequisite-only path recommendation by generating dual knowledge structures automatically and using similarity-based fallback during learning. The framework is general with respect to datasets that provide KC names, and the use of TextGrad to refine LLM-generated explanations is a concrete methodological addition. The authors compare against a broad set of baselines, include an ablation isolating the similarity agent, and attempt significance testing. However, because the evaluation metric is computed from the same knowledge-tracing model that provides the RL reward and is trained on the full dataset, the central SOTA claim is not yet established; the manuscript also does not validate the generated graphs against expert structure. These are fixable with additional experiments, so the work is a strong candidate after revision.

major comments (4)
  1. [Section 5.3 and Eq. (1)/(13)] The evaluation quantity Ep is computed from DKT-predicted mastery, and the same Ep is used as the terminal RL reward in Eq. (13). Moreover, Section 5.3 states that the DKT evaluator is trained on the entire dataset, so the evaluation is not a held-out estimate. The policy is thus directly optimized to maximize the exact quantity reported in Table 1, and the gains may partly reflect idiosyncrasies of this particular KT model. Please provide a concrete external check: train the evaluator on a disjoint portion of the logs, evaluate with an alternative KT model, or report post-test style mastery outcomes; without such a check the SOTA claim is not convincing.
  2. [Table 1] No variance or number of seeds is reported, and the p-value notation is not backed by a description of the test or the comparison distribution. For example, the Junyi step-5 margin of KnowLP over DLPR is 0.2406 vs 0.2086, but without error bars one cannot tell whether this gap is within run-to-run noise. Please report mean and standard deviation over at least five random seeds and state the exact statistical test used.
  3. [Section 5.6] The generated KC graphs are analyzed only via edge counts and visualization; there is no quantitative comparison against expert-annotated prerequisite relations (available for Junyi) or against any human judgment of the generated similarity edges. Since the paper's two main components—EDU-GraphRAG and the similarity agent—depend on these graphs, the reader cannot determine whether the performance gains come from accurate structure or from spurious edges that happen to improve the in-loop metric. Please add precision/recall against the available Junyi annotations and a sample-based expert evaluation of the similarity edges on all datasets.
  4. [Section 5.8] The KES simulator is built on a DKT model, the same model family used for the reward and the evaluation, and the paper explicitly states that real-world static data cannot validate effectiveness. As it stands, the simulation only shows that KnowLP performs well when the student is modeled by the same proxy used to train the policy. This does not break the circularity. Please either include a human-subject or field evaluation, or simulate students with a model that is not derived from the reward/evaluation DKT, and clearly report the simulation as a sanity check rather than as evidence of real-world effectiveness.
minor comments (8)
  1. [Section 5.7] Section 5.7 is a verbatim repetition of Section 5.5; the actual case-study content (Figure 5) is not described, so the interpretability claim is not supported by the text.
  2. [Table 3] In Table 3, the row labels list step=15 twice; the fourth row should presumably be step=20.
  3. [Eq. (14)] Eq. (14) uses the notation difficult(e_ci^m) without defining the superscript m; clarify that it is the difficulty of the m-th exercise associated with KC c_i.
  4. [Section 4.2.5, Eq. (15)] Eq. (15) defines p_t using h^{c_t}_{t-1} and h^{c_t}_{t-2}, but the indexing of the KC under consideration is ambiguous when the learner has not practiced the same KC in consecutive steps; clarify the notation.
  5. [Eqs. (11)-(12)] The PPO objective in Eq. (12) and the advantage in Eq. (11) are written in a non-standard form; please reconcile with a standard reference or define \hat{A}_t more carefully.
  6. [Sections 5.2-5.3] The paper uses 'DKT' both for the evaluation model and for a baseline family; please disambiguate the terminology in Section 5.2 and 5.3.
  7. [Figure 3] Figure 3 is difficult to read at the current resolution; please use larger panels and clearer labels, and report the edge counts in a table as well.
  8. [Section 4.2.2] The reference to GEHRL ('optimize PPO using a loss function similar to GEHRL') appears to cite [3] (SRC) rather than [21]; please correct the citation.

Circularity Check

2 steps flagged · score 6.0 of 10

The SOTA result is measured on the same DKT-predicted improvement that the RL reward maximizes (Eq. 13 = Eq. 1), with the evaluator trained on the full dataset, so the headline gain is partially circular.

  1. fitted input called prediction [Eq. (1), Eq. (13), Section 5.3]
    "To maximize the improvement in learners’ mastery of the learning objectives, we calculate the corresponding reward value after selecting knowledge concepts at each step. R_et = (Ep, if t is the last learning stage; 0, otherwise.) (13) ... To more accurately evaluate the model’s performance, we trained a DKT [29] on the entire dataset to predict learners’ mastery of target knowledge concepts in Eq. (1)."

    The reported metric Ep in Eq. (1) is the normalized gain in DKT-predicted mastery, and Section 5.3 states that this DKT is trained on the entire dataset. Eq. (13) sets the terminal reinforcement-learning reward to exactly Ep. Therefore the quantity in Table 1 that supports the state-of-the-art claim is the same quantity the policy is trained to maximize; the evaluation is not an independent outcome measure. A policy can inflate Ep by exploiting DKT’s predictions, and the paper provides no external validation that would break this identity.

  2. other [Section 5.8]
    "Therefore, we used a Knowledge Evolution-based Simulator (KES), which leverages the DKT model [29] to simulate students’ exercise behavior on randomly generated exercise sequences."

    The paper’s only attempt to validate effectiveness in online scenarios uses a simulator built on DKT, the same knowledge-tracing model family used to compute Ep and to reward the policy. The simulation therefore reuses the in-loop predictor rather than testing against real learning outcomes, so it does not provide independent evidence that the optimized Ep corresponds to genuine learning gains.

full rationale

The graph-generation and multi-agent parts of KnowLP are not equation-level circular: EDU-GraphRAG is an LLM pipeline whose outputs are not defined in terms of the evaluation metric, and no load-bearing self-citation chain is used. The circularity is concentrated in the evaluation loop. Eq. (13) makes the reported improvement Ep of Eq. (1) the exact RL reward, and Section 5.3 says the DKT used for Ep is trained on the entire dataset, so the headline improvements over DLPR and other baselines are measured on the objective being optimized. Section 5.8 strengthens this concern by replacing real-world validation with a DKT-based simulator. This is not a case of a fitted constant being renamed as a prediction; it is an in-loop evaluator/reward identity that makes the central SOTA claim partially circular. If a held-out real-outcome test or an expert-validated graph comparison were provided, the claim would have independent support.

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

The central claim rests on LLM-generated graph quality, the DKT/DIMKT proxy for learning, a hand-set activation threshold, and per-dataset chunk sizes. No new physical entities are introduced. The paper does not provide an external validation of the generated graph against expert annotations or of the recommendation against real learning outcomes.

free parameters (4)
  • S-agent activation threshold tau = 0.001
    Set based on the observed DKT average mastery improvement per exercise (§5.3); used to decide when the similarity agent fires (§4.2.5).
  • EDU-GraphRAG chunk sizes = Junyi 800, ASS09 600, MOOCCubeX 400
    Chosen per dataset for text segmentation (§5.3); no sensitivity analysis or search procedure reported.
  • DKT mastery threshold = 0.5
    Mastery criterion used in Eq. (1); adopted from prior work without validation on these datasets.
  • PPO hyperparameters = epsilon=0.2, learning rate=0.001
    Standard PPO defaults stated in §5.3; no tuning or sensitivity analysis reported.
assumptions (4)
  • domain assumption LLM-generated prerequisite and similarity edges derived from KC names and generated explanations are sufficiently accurate and complete to guide learning path recommendation.
    EDU-GraphRAG (§4.1) produces the graphs without expert validation; correctness is asserted through downstream performance, not measured directly.
  • domain assumption DKT/DIMKT predicted mastery is a valid proxy for real learning gain.
    Eq. (1) and §4.2.1 use predicted knowledge states for reward and evaluation; no external post-test or human study is provided.
  • domain assumption Discrimination learning (interleaving similar concepts) improves learning in this setting.
    Borrowed from Gagne's learning hierarchy and interleaving literature; assumed to transfer to learning path recommendation without direct evidence here.
  • ad hoc to paper The hand-set S-agent trigger threshold tau=0.001 captures the notion of learning blockage.
    §4.2.5 defines blockage as improvement below tau; tau is fit to DKT average improvement, not derived from learner behavior or learning theory.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GraphRAG-Induced Dual Knowledge Structure Graphs for Personalized Learning Path Recommendation." pith.science (2026). https://pith.science/paper/BMHNU4VC

@misc{pith2026250622303,
  author       = {Pith},
  title        = {Pith review of: GraphRAG-Induced Dual Knowledge Structure Graphs for Personalized Learning Path Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BMHNU4VC}},
  note         = {Machine review of arXiv:2506.22303}
}
read the original abstract

Learning path recommendation seeks to provide learners with a structured sequence of learning items (\eg, knowledge concepts or exercises) to optimize their learning efficiency. Despite significant efforts in this area, most existing methods primarily rely on prerequisite relationships, which present two major limitations: 1) Requiring prerequisite relationships between knowledge concepts, which are difficult to obtain due to the cost of expert annotation, hindering the application of current learning path recommendation methods. 2) Relying on a single, sequentially dependent knowledge structure based on prerequisite relationships implies that difficulties at any stage can cause learning blockages, which in turn disrupt subsequent learning processes. To address these challenges, we propose a novel approach, GraphRAG-Induced Dual Knowledge Structure Graphs for Personalized Learning Path Recommendation (KnowLP), which enhances learning path recommendations by incorporating both prerequisite and similarity relationships between knowledge concepts. Specifically, we introduce a knowledge concept structure graph generation module EDU-GraphRAG that adaptively constructs knowledge concept structure graphs for different educational datasets, significantly improving the generalizability of learning path recommendation methods. We then propose a Discrimination Learning-driven Reinforcement Learning (DLRL) module, which mitigates the issue of blocked learning paths, further enhancing the efficacy of learning path recommendations. Finally, we conduct extensive experiments on three benchmark datasets, demonstrating that our method not only achieves state-of-the-art performance but also provides interpretable reasoning for the recommended learning paths.

Figures

Figures reproduced from arXiv: 2506.22303 by the authors.

Figure 1
Figure 1. The comparison of traditional methods (top) and our proposed KnowLP (bottom) is depicted. The top denotes [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. KnowLP Framework Overview. The KC structure graph generation module on the left begins by inputting KC names [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Comparison between the original graph and the [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: The explanation of learning path [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: Results of simulation experiment. 5.8 Simulation Experiment Since real-world data only contains static information, which can￾not be directly used to analyze the exercise sequences of students not included in the dataset, the effectiveness of our method in real online …

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. LiveGraph: Active-Structure Neural Re-ranking for Exercise Recommendation

    cs.IR 2026-02 unverdicted novelty 5.0 of 10

    LiveGraph re-ranks exercise recommendations with a dynamic concept kernel and uncertainty-aware meta-RL fusion, claiming improved accuracy and diversity, with only accuracy tables shown.

Reference graph

Works this paper leans on

49 extracted references · 38 canonical work pages · cited by 1 Pith paper

  1. [1]

    Xiaoyan Cai, Junwei Han, Wenjie Li, Renxian Zhang, Shirui Pan, and Libin Yang. 2018. A three-layered mutually reinforced model for personalized citation recommendation. IEEE transactions on neural networks and learning systems 29, 12 (2018), 6026–6037

    work page 2018

  2. [2]

    Chih-Ming Chen. 2009. Ontology-based concept map for planning a personalised learning path. British Journal of Educational Technology 40, 6 (2009), 1028–1058

    work page 2009

  3. [3]

    Xianyu Chen, Jian Shen, Wei Xia, Jiarui Jin, Yakun Song, Weinan Zhang, Weiwen Liu, Menghui Zhu, Ruiming Tang, Kai Dong, et al. 2023. Set-to-sequence Ranking- based Concept-aware Learning Path Recommendation. In Proceedings of 2023 AAAI Conference on Artificial Intelligence (AAAI-2023) . 5027–5035

    work page 2023

  4. [4]

    Xinghe Cheng, Xufang Zhou, Liangda Fang, Chaobo He, Yuyu Zhou, Weiqi Luo, Zhiguo Gong, and Quanlong Guan. 2025. NR4DER: Neural Re-ranking for Diversified Exercise Recommendation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR-2025). 1738–1747

    work page 2025

  5. [5]

    WOROKWU Chinda. 2022. Effect of Gagne’s learning hierarchy on Chemistry Student achievement in senior secondary school. Iconic Research and Engineering Journals 5, 9 (2022), 609–618

    work page 2022

  6. [6]

    Thomas Cover and Peter Hart. 1967. Nearest neighbor pattern classification. IEEE transactions on information theory 13, 1 (1967), 21–27

    1967

  7. [7]

    Kingma Diederik and Jimmy Lei Ba

    P. Kingma Diederik and Jimmy Lei Ba. 2015. Adam: A Method for Stochastic Optimization. In Proceedings of the 3rd International Conference for Learning Representations (ICLR-2015)

    work page 2015

  8. [8]

    František Duchoň, Andrej Babinec, Martin Kajan, Peter Beňo, Martin Florek, Tomáš Fico, and Ladislav Jurišica. 2014. Path planning with modified a star algorithm for a mobile robot. Procedia engineering 96 (2014), 59–69

    work page 2014

Show all 49 references
  1. [9]

    Pragya Dwivedi, Vibhor Kant, and Kamal K Bharadwaj. 2018. Learning path recommendation based on modified variable length genetic algorithm. Education and information technologies 23 (2018), 819–836

    2018

  2. [10]

    Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, and Jonathan Larson. 2024. From local to global: A graph rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130 (2024)

    2024 arXiv

  3. [11]

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997 (2023)

    2023 arXiv

  4. [12]

    Travis R Goodwin, Max E Savery, and Dina Demner-Fushman. 2020. Flight of the PEGASUS? Comparing Transformers on Few-shot and Zero-shot Multi-document Abstractive Summarization. In Proceedings of 2020 COLING. international confer- ence on computational linguistics (COLING-2020)...

    2020

  5. [13]

    Quanlong Guan, Xinghe Cheng, Fang Xiao, Zhuzhou Li, Chaobo He, Liangda Fang, Guanliang Chen, Zhiguo Gong, and Weiqi Luo. 2025. Explainable exercise recommendation with knowledge graph. Neural Networks 183 (2025), 106954

    2025

  6. [14]

    B Hidasi. 2015. Session-based Recommendations with Recurrent Neural Networks. arXiv preprint arXiv:1511.06939 (2015)

    2015 arXiv

  7. [15]

    Wacharawan Intayoad, Chayapol Kamyod, and Punnarumol Temdee. 2020. Rein- forcement learning based on contextual bandits for personalized online learning recommendation systems. Wireless Personal Communications 115, 4 (2020), 2917– 2932

    2020

  8. [16]

    Vijay Konda and John Tsitsiklis. 1999. Actor-critic algorithms.Advances in neural information processing systems 12 (1999)

    1999

  9. [17]

    Yoshiki Kubotani, Yoshihiro Fukuhara, and Shigeo Morishima. 2021. Rltutor: Reinforcement learning based adaptive tutoring system by modeling virtual student with fewer interactions. arXiv preprint arXiv:2108.00268 (2021)

    2021 arXiv

  10. [18]

    Md Tahmid Rahman Laskar, Enamul Hoque, and Jimmy Huang. 2020. Query Focused Abstractive Summarization via Incorporating Query Relevance and Transfer Learning with Transformer Models. In Proceedings of 2020 Advances in Artificial Intelligence - 33rd Canadian Conference on Artif...

    2020

  11. [19]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al

  12. [20]

    Qingyao Li, Wei Xia, Li’ang Yin, Jiarui Jin, and Yong Yu. 2024. Privileged Knowl- edge State Distillation for Reinforcement Learning-based Educational Path Rec- ommendation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD-2024) . 1621–1630

    2024

  13. [21]

    Qingyao Li, Wei Xia, Li’ang Yin, Jian Shen, Renting Rui, Weinan Zhang, Xianyu Chen, Ruiming Tang, and Yong Yu. 2023. Graph Enhanced Hierarchical Reinforce- ment Learning for Goal-oriented Learning Path Recommendation. In Proceedings of the 32nd ACM International Conference on ...

    2023

  14. [22]

    Chun Fu Lin, Yu-chu Yeh, Yu Hsin Hung, and Ray I Chang. 2013. Data mining for providing a personalized learning path in creativity: An application of decision trees. Computers & Education 68 (2013), 199–210

    2013

  15. [23]

    Qi Liu, Shiwei Tong, Chuanren Liu, Hongke Zhao, Enhong Chen, Haiping Ma, and Shijin Wang. 2019. Exploiting Cognitive Structure for Adaptive Learning. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining (KDD-2019) . 627–635

    2019

  16. [24]

    Sein Minn, Feida Zhu, and Michel C Desmarais. 2018. Improving Knowledge Tracing Model by Integrating Problem Difficulty. In Proceedings of 2018 IEEE International conference on data mining workshops (ICDMW-2018) . IEEE, 1505– 1506

    2018

  17. [25]

    Amir Hossein Nabizadeh, Daniel Goncalves, Sandra Gama, Joaquim Jorge, and Hamed N Rafsanjani. 2020. Adaptive learning path recommender approach using auxiliary learning objects. Computers & Education 147 (2020), 103777

    2020

  18. [26]

    Amir Hossein Nabizadeh, Alípio Mário Jorge, and José Paulo Leal. 2017. Rutico: Recommending Successful Learning Paths under Time Constraints. InProceedings of the 25th conference on user modeling, adaptation and personalization (UMAP- 2017). 153–158

    2017

  19. [27]

    Steven C Pan, Ganeash Selvarajan, and Chanda S Murphy. 2024. Interleaved pretesting enhances category learning and classification skills. Journal of Applied Research in Memory and Cognition 13, 3 (2024), 393

    2024

  20. [28]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al

  21. [29]

    Chris Piech, Jonathan Bassen, Jonathan Huang, Surya Ganguli, Mehran Sahami, Leonidas J Guibas, and Jascha Sohl-Dickstein. 2015. Deep Knowledge Tracing. Proceedings of the 28th International Conference on Neural Information Processing Systems (NIPS-2015), 505–513

    2015

  22. [30]

    Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. 2023. In-context retrieval-augmented language models. Transactions of the Association for Computational Linguistics 11 (2023), 1316–1331

    2023

  23. [31]

    Doug Rohrer. 2012. Interleaving helps students distinguish among similar con- cepts. Educational Psychology Review 24 (2012), 355–367

    2012

  24. [32]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov

  25. [33]

    Shuanghong Shen, Zhenya Huang, Qi Liu, Yu Su, Shijin Wang, and Enhong Chen

  26. [34]

    Daqian Shi, Ting Wang, Hao Xing, and Hao Xu. 2020. A learning path recom- mendation model based on a multidimensional knowledge graph framework for e-learning. Knowledge-Based Systems 195 (2020), 105618

    2020

  27. [35]

    Hangyu Wang, Ting Long, Liang Yin, Weinan Zhang, Wei Xia, Qichen Hong, Dingyin Xia, Ruiming Tang, and Yong Yu. 2023. GMOCAT: A Graph-Enhanced Multi-Objective Method for Computerized Adaptive Testing. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and D...

    2023

  28. [36]

    Jiapu Wang, Sun Kai, Linhao Luo, Wei Wei, Yongli Hu, Alan Wee-Chung Liew, Shirui Pan, and Baocai Yin. 2024. Large language models-guided dynamic adap- tation for temporal knowledge graph reasoning. Advances in Neural Information Processing Systems 37 (2024), 8384–8410

    2024

  29. [37]

    Dihua Xu, Zhijian Wang, Kejia Chen, and Weidong Huang. 2012. Personalized Learning Path Recommender Based on User Profile Using Social Tags. In Pro- ceedings of the 5th International Symposium on Computational Intelligence and Design (ISCID-2012), Vol. 1. IEEE, 511–514

    2012

  30. [38]

    Xiaocheng Yang, Mingyu Yan, Shirui Pan, Xiaochun Ye, and Dongrui Fan. 2023. Simple and Efficient Heterogeneous Graph Neural Network. In Proceedings of 2023 AAAI conference on artificial intelligence (AAAI-2023) . 10816–10824

    2023

  31. [39]

    Differentiation

    Mert Yuksekgonul, Federico Bianchi, Joseph Boen, Sheng Liu, Zhi Huang, Carlos Guestrin, and James Zou. 2024. TextGrad: Automatic" Differentiation" via Text. arXiv preprint arXiv:2406.07496 (2024)

    2024 arXiv

  32. [40]

    Mert Yuksekgonul, Federico Bianchi, Joseph Boen, Sheng Liu, Pan Lu, Zhi Huang, Carlos Guestrin, and James Zou. 2025. Optimizing generative AI by backpropa- gating language model feedback. Nature 639 (2025), 609–616

    2025

  33. [41]

    Haotian Zhang, Chenyang Bu, Fei Liu, Shuochen Liu, Yuhong Zhang, and Xue- gang Hu. 2022. APGKT: Exploiting Associative Path on Skills Graph for Knowl- edge Tracing. In Proceedings of 2022 Pacific Rim International Conference on Arti- ficial Intelligence (PRICAI-2022). Lecture ...

    2022

  34. [42]

    Haotian Zhang, Shuanghong Shen, Bihan Xu, Zhenya Huang, Jinze Wu, Jing Sha, and Shijin Wang. 2024. Item-Difficulty-Aware Learning Path Recommenda- tion: From a Real Walking Perspective. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KD...

    2024

  35. [43]

    Lei Zhang, Wuji Zhang, Likang Wu, Ming He, and Hongke Zhao. 2023. SHGCN: Socially enhanced heterogeneous graph convolutional network for multi- behavior prediction. ACM Transactions on the Web 18, 1 (2023), 1–27

    2023

  36. [44]

    Qian Zhang, Jie Lu, and Guangquan Zhang. 2021. Recommender Systems in E-learning. Journal of Smart Environments and Green Computing 1, 2 (2021), 76–89

    2021

  37. [45]

    Chuanpan Zheng, Xiaoliang Fan, Shirui Pan, Haibing Jin, Zhaopeng Peng, Zong- han Wu, Cheng Wang, and S Yu Philip. 2023. Spatio-temporal joint graph con- volutional networks for traffic forecasting. IEEE Transactions on Knowledge and Data Engineering 36, 1 (2023), 372–385

    2023

  38. [2017]

    arXiv preprint arXiv:1707.06347 (2017)

    Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)

    2017 arXiv

  39. [2019]

    In Proceedings of 2019 Advances in Neural Information Processing Systems (NeurIPS- 2019)

    Pytorch: an Imperative Style, High-performance Deep Learning Library. In Proceedings of 2019 Advances in Neural Information Processing Systems (NeurIPS- 2019). 8024–8035

    2019

  40. [2020]

    In Proceedings of 2020 Advances in Neural Information Processing Systems (NeurIPS- 2020)

    Retrieval-augmented Generation for Knowledge-intensive Nlp Tasks. In Proceedings of 2020 Advances in Neural Information Processing Systems (NeurIPS- 2020). 9459–9474. Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Xinghe Cheng et al

    2020

  41. [2022]

    In Proceedings of the 45th international ACM SIGIR conference on research and development in information retrieval (SIGIR-2022)

    Assessing Student’s Dynamic Knowledge State by Exploring the Question Difficulty Effect. In Proceedings of the 45th international ACM SIGIR conference on research and development in information retrieval (SIGIR-2022) . 427–437

    2022

Pith tools

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