Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

LLM-KT: Aligning Large Language Models with Knowledge Tracing using a Plug-and-Play Instruction

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read LLM-KT achieves state-of-the-art knowledge tracing by feeding compressed question context and ID-sequence embeddings into a large language model.

desk verdict Sensible LLM-plus-sequence-model architecture with clean internal ablations, but the SOTA claim is built on cross-paper baseline numbers under unknown splits and needs a shared-protocol re-run before it is load-bearing. read the letter →

arxiv 2502.02945 v1 pith:MXWFHCPB submitted 2025-02-05 cs.CL cs.AI

Ziwei Wang , Jie Zhou , Qin Chen , Min Zhang , Bo Jiang , Aimin Zhou , Qinchun Bai , Liang He This is my paper · ORCID
classification cs.CLcs.AI
keywords knowledgetracinglargelanguagemodelsplug-and-playinstructionplug-incontextsequenceLoRAstudentperformancepredictionpersonalizededucation
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

Knowledge tracing is the task of predicting whether a student will answer the next question correctly from their past question-answer record, and it underpins adaptive tutoring systems. The paper aims to establish that the two existing model families are each half right: sequence-based models capture interaction patterns from question and concept IDs, while large language models can reason with the world knowledge in question text, but neither alone handles both well. LLM-KT is the proposed reconciliation: a plug-and-play instruction template aligns an LLM with the tracing task, letting one model combine both strengths. The paper reports that this framework outperforms roughly twenty strong baselines across four standard datasets, with the largest gains coming from injecting both a compressed textual context and traditional sequence-model embeddings into the LLM. If the claim holds, LLMs become a practical backbone for personalized education rather than a text-only reasoning tool.

What carries the argument

The load-bearing object is the Plug-and-Play Instruction: a prompt template with question-specific tokens ([QuesEmbed...]) and concept-specific tokens ([ConcEmbed...]) that act as insertion slots for two aligned representations. The plug-in context module encodes question and concept text with a frozen encoder (LLaMA2 in the main configuration) and maps it through an MLP context adapter into the LLM embedding space; the plug-in sequence module takes ID embeddings produced by a traditional sequence model (AKT in the main configuration) and maps them through a sequence adapter. The two representations are combined by addition, and the resulting token embeddings are inserted into the LLM, which is fine-tuned with LoRA. This carries the argument by making the LLM's reasoning operate on the same compressed objects that sequence models use, rather than on raw long texts or raw IDs.

What would settle it

Re-run every baseline on the identical 8:1:1 student split and preprocessing used by LLM-KT, measure AUC and ACC on the same test partition, and check the paired differences; if most of the twenty comparisons stop favoring LLM-KT, the state-of-the-art claim is refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that knowledge tracing can be reformulated as an instruction-following task for a large language model, and that this reformulation succeeds only when the LLM receives both kinds of information it lacks: compressed representations of long question and concept texts (the plug-in context) and interaction-aware representations of the ID sequence from a traditional sequence model (the plug-in sequence). In LLM-KT, a student's history is written as a natural-language prompt in which each question and concept carries a special token whose embedding is the sum of a text-derived embedding and an ID-sequence-derived embedding, and the LLM is fine-tuned with LoRA to answer 'Yes' or 'No' for the target question. On the paper's reported experiments, this design achieves the best AUC and ACC on Assist2009, Assist2015, Junyi, and Nips2020, with AUC gains over the strongest LLM-only baseline ranging from roughly 2 to 5.7 points depending on the dataset.

Load-bearing premise

The comparison with roughly twenty baselines assumes those baselines' reported numbers were produced under the same data split and preprocessing as LLM-KT's own 8:1:1 split, but the paper takes baseline numbers from earlier papers while running its own model on its own split.

Editorial extensions

If this is right

  • If the central claim holds, an LLM can serve as the backbone for knowledge tracing while a traditional sequence model remains a plug-in, so future tracing systems need not choose between behavior modeling and semantic reasoning.
  • The plug-in context design implies that long question histories can be compressed into a handful of token embeddings without losing the semantic signal, and the paper shows improving results as sequence length grows to 100 in most datasets.
  • The ablation results imply that both modalities carry complementary information: removing either the plug-in context or the plug-in sequence lowers AUC on the two datasets where both are available.
  • The comparison with LLM-FTTokenID implies that initializing question and concept token embeddings from DKT or AKT is much more effective than learning them from scratch during LLM fine-tuning.

Reading between the lines

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

  • The paper leaves implicit that the same instruction-with-special-tokens pattern could transfer to other sequential education predictions, such as hint usefulness or dropout, where item text and interaction order both matter.
  • Because question text is encoded offline by a frozen encoder, the approach should extend to cold-start questions by deriving their embeddings from text alone, although LLM-KT does not report such an experiment.
  • A natural testable extension would replace the fixed addition operation for merging context and sequence embeddings with a learned gating mechanism, which might sharpen the interaction between the two modalities.
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

3 major / 5 minor

Summary. The paper proposes LLM-KT, a knowledge tracing framework that fine-tunes LLaMA2 (via LoRA) with a plug-and-play instruction containing special tokens for question IDs, concept IDs, and compressed text representations. A Plug-in Context module encodes question/concept text with a separate encoder and adapts it into the LLM embedding space, while a Plug-in Sequence module injects ID-sequence representations learned by a traditional sequence model (AKT or DKT). The model is evaluated on ASSIST2009, ASSIST2015, Junyi, and NIPS2020, reporting AUC and ACC against roughly 20 baselines, with ablations and analyses of sequence length and encoder choices.

Significance. If the central state-of-the-art claim were established under a common protocol, the paper would make a useful contribution: it offers a modular and relatively clean way to combine LLM text understanding with traditional sequence-interaction models, and the component ablations and encoder analyses provide useful evidence about the contribution of each plug-in. The appendix prompt templates are detailed and the sequence-length diagnostics across four datasets are a helpful addition. However, the significance is currently limited by the evaluation protocol: the headline comparison uses baseline numbers borrowed from heterogeneous sources, the NIPS2020 dataset is modified in a non-released way, and no error bars or significance tests are reported.

major comments (3)
  1. [§4.2–§4.3 and Table 2] The central claim that LLM-KT achieves state-of-the-art results is not established by the comparison as presented. Section 4.3 states that all experiments use an 8:1:1 train/validation/test split following MRT-KT, while Section 4.2 states that DL-, PLM-, and context-aware baseline numbers are taken from the original papers or other relevant papers. Those papers do not necessarily use the same split, student filtering, sequence truncation, or dataset preprocessing (e.g., Junyi and NIPS2020 are commonly processed differently across publications). Without re-running the baselines under the identical protocol, the AUC/ACC differences in Table 2—some as small as +0.007 on Junyi AUC—cannot be attributed to the method rather than to protocol mismatch. I would need to see a common-protocol re-run of at least the strongest baselines (AKT, LPKT, LBKT, MRT-KT) to support the SOTA claim.
  2. [§4.1.1 and Table 1] The NIPS2020 results are not reproducible as reported. The dataset is modified by manually converting figures to text, and this converted version is not released. In addition, the text says “the top 150 most frequently appearing questions” while Table 1 lists 110 questions for NIPS2020; this discrepancy must be resolved. Because the NIPS2020 SOTA claim rests on a non-public, apparently custom preprocessing, readers cannot verify or build on these numbers.
  3. [§5.1, Table 2] No error bars or significance tests are reported for any method, and several claimed advantages are small or go in the opposite direction on one metric. For example, on Junyi, LLM-KT’s ACC (0.8294) is below LPKT (0.8344) and LBKT† (0.8362) in the same table, so the abstract’s phrase “state-of-the-art performance on four typical datasets” is too strong even before accounting for protocol differences. At minimum, the paper should report standard deviations over multiple seeds and use a significance test, or clearly state which metrics are not SOTA.
minor comments (5)
  1. [§4.1.1 and Table 1] Assist2015 is described as containing “100 distinct questions” but Table 1 shows no question count; please clarify whether question IDs exist in the version used or are discarded.
  2. [§5.4.1] The phrase “we calculate the the vector” contains a duplicated article and should be corrected; also, the recommendation to use MPNET is not supported by the reported results, since LLaMA2 achieves the highest average AUC and ACC in the same table.
  3. [§3.3 and §4.3] The paper should state explicitly whether the sequence encoder (AKT or DKT) is trained jointly with the LLM or pretrained and frozen, and should report its training details (e.g., loss, epochs, and whether the same 8:1:1 split is used for its training).
  4. [Table 2 caption and §4.2] The source of each baseline number should be given per cell or at least per row, since the caption already notes that some rows are filled from other papers; a reader cannot otherwise tell which numbers come with a matching protocol.
  5. [§4.2.4, Table 2] The GPT-4o results are only reported for two datasets; please specify the prompt template, decoding settings, and the exact model version used so those numbers are reproducible.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: LLM-KT is a standard supervised training/evaluation pipeline; the only self-referential element, following MRT-KT for the data split, is non-load-bearing.

full rationale

LLM-KT is trained with cross-entropy on labeled KT records and evaluated on held-out test records (Section 4.3: 'we divide the student dataset in a ratio of 8:1:1 for training, validation, and test sets'). Equations (1)-(6) construct question/concept embeddings from text and ID encoders, and Eq. (2) computes P(Yes|x) from the LLM's own Yes/No logits; no target quantity is defined in terms of the model's output, and no parameter is fitted on the test set. The only self-referential element is the citation of MRT-KT [6] for the 8:1:1 split; since Aimin Zhou is a co-author of both papers, this is a self-citation, but it is a protocol choice rather than a fitted parameter or a derivation step, so it does not force the comparison. The SOTA claim rests on cross-paper baseline numbers taken from original papers (Section 4.2: 'we only report the results from the original papers or other relevant experimental papers'), raising a protocol-mismatch correctness risk, but that is an experimental validity concern, not a circular derivation. No uniqueness theorem, imported ansatz, or renamed empirical pattern is load-bearing. Under the standard supervised train/validate/test paradigm the derivation is self-contained.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper's central claim depends on standard KT task assumptions, a particular data split, a manual and unreleased data conversion, and the assertion that static ID embeddings from AKT represent sequence interaction behavior. None of these are independently verified in the paper.

free parameters (6)
  • Sequence length of history records = 100
    Selected in implementation (Section 4.3), with length sweeps in Section 5.3; affects LLM prompt size.
  • LoRA rank/alpha/dropout = 32/32/0.1
    Implementation hyperparameters (Section 4.3) chosen without ablation.
  • Learning rate and weight decay = 3e-4 / 1e-5
    Implementation hyperparameters (Section 4.3).
  • Combination function g = addition
    Chosen by validation comparison in Section 5.4.3.
  • Sequence encoder choice = AKT
    Selected by validation over DKT and AKT in Section 5.4.2.
  • Context encoder choice = LLaMA2
    Selected by validation over BERT, MPNET, and LLaMA2 in Section 5.4.1.
assumptions (5)
  • domain assumption A student's future performance is predictable from the chronological sequence of question IDs, concept IDs, and question text.
    This is the standard knowledge tracing premise, invoked throughout Section 3 as the task definition.
  • domain assumption The 8:1:1 random split used in Section 4.3 yields an unbiased evaluation comparable to the splits used when baseline numbers were reported.
    The paper follows MRT-KT's split but does not verify that borrowed baseline results used the same split.
  • ad hoc to paper Manually converting Nips2020 figures into text preserves the information needed for prediction.
    The conversion is a manual, unverified step described in Section 4.1.1.
  • domain assumption LLMs have world knowledge and reasoning capacity that improve knowledge tracing.
    This premises the entire LLM-based approach, stated in the introduction.
  • ad hoc to paper Static question ID embeddings trained by AKT/DKT encode the sequence interaction behavior that the LLM needs.
    In Eq. 5 the sequence encoder takes only a question ID and returns a fixed embedding; the claim that this captures interaction behavior is asserted, not derived.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLM-KT: Aligning Large Language Models with Knowledge Tracing using a Plug-and-Play Instruction." pith.science (2026). https://pith.science/paper/MXWFHCPB

@misc{pith2026250202945,
  author       = {Pith},
  title        = {Pith review of: LLM-KT: Aligning Large Language Models with Knowledge Tracing using a Plug-and-Play Instruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MXWFHCPB}},
  note         = {Machine review of arXiv:2502.02945}
}
read the original abstract

The knowledge tracing (KT) problem is an extremely important topic in personalized education, which aims to predict whether students can correctly answer the next question based on their past question-answer records. Prior work on this task mainly focused on learning the sequence of behaviors based on the IDs or textual information. However, these studies usually fail to capture students' sufficient behavioral patterns without reasoning with rich world knowledge about questions. In this paper, we propose a large language models (LLMs)-based framework for KT, named \texttt{\textbf{LLM-KT}}, to integrate the strengths of LLMs and traditional sequence interaction models. For task-level alignment, we design Plug-and-Play instruction to align LLMs with KT, leveraging LLMs' rich knowledge and powerful reasoning capacity. For modality-level alignment, we design the plug-in context and sequence to integrate multiple modalities learned by traditional methods. To capture the long context of history records, we present a plug-in context to flexibly insert the compressed context embedding into LLMs using question-specific and concept-specific tokens. Furthermore, we introduce a plug-in sequence to enhance LLMs with sequence interaction behavior representation learned by traditional sequence models using a sequence adapter. Extensive experiments show that \texttt{\textbf{LLM-KT}} obtains state-of-the-art performance on four typical datasets by comparing it with approximately 20 strong baselines.

Figures

Figures reproduced from arXiv: 2502.02945 by the authors.

Figure 1
Figure 1. The advantages of traditional models and LLMs for [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The framework of LLM-KT. We propose a Plug-and-Play Instruction to combine the strengths of LLMs and traditional sequence models for knowledge tracing by inserting multiple modalities into LLMs. Particularly, we design a Plug-in Context module to capture the long context of students’ problem-solving records. Then, we introduce the Plug-in Sequence to align the sequence interaction representation learned by the tradi… view at source ↗
Figure 3
Figure 3. Influence of sequence length on four different datasets [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Influence of sequence length on four different datasets [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Continue with ORCID 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. Constructing a Question-Answering Simulator through the Distillation of LLMs

    cs.LG 2025-09 conditional novelty 6.0 of 10

    LDSim distills an LLM's concept-prerequisite knowledge and mastery reasoning into a lightweight simulator that beats LLM-based and LLM-free baselines on four knowledge-tracing datasets.

Reference graph

Works this paper leans on

45 extracted references · 31 canonical work pages · cited by 1 Pith paper

  1. [1]

    Ghodai Abdelrahman, Qing Wang, and Bernardo Nunes. 2023. Knowledge tracing: A survey. Comput. Surveys 55, 11 (2023), 1–37

    2023

  2. [2]

    bigdata ustc. 2021. EduData. https://github.com/bigdata-ustc/EduData

    work page 2021

  3. [3]

    Song Cheng, Qi Liu, Enhong Chen, Kai Zhang, Zhenya Huang, Yu Yin, Xiaoqing Huang, and Yu Su. 2022. AdaptKT: A domain adaptable method for knowledge tracing. In Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining. 123–131

    work page 2022

  4. [4]

    Corbett and John R

    Albert T. Corbett and John R. Anderson. 1995. Knowledge tracing: Modeling the acquisition of procedural knowledge. User Modelling and User-Adapted Interaction (Jan 1995), 253–278. https://doi.org/10.1007/bf01099821

    doi:10.1007/bf01099821 1995

  5. [5]

    Chaoran Cui, Yumo Yao, Chunyun Zhang, Hebo Ma, Yuling Ma, Zhaochun Ren, Chen Zhang, and James Ko. 2024. DGEKT: A Dual Graph Ensemble Learning Method for Knowledge Tracing. ACM Trans. Inf. Syst. 42, 3, Article 78 (Jan. 2024), 24 pages. https://doi.org/10.1145/3638350

    doi:10.1145/3638350 2024

  6. [6]

    Jiajun Cui, Zeyuan Chen, Aimin Zhou, Jianyong Wang, and Wei Zhang. 2023. Fine-grained interaction modeling with multi-relational transformer for knowledge tracing. ACM Transactions on Information Systems 41, 4 (2023), 1–26

    work page 2023

  7. [7]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In North American Chapter of the Association for Computational Linguistics . https://api.semanticscholar.org/CorpusID:52967399

    work page 2019

  8. [8]

    Jeffrey L Elman. 1990. Finding structure in time. Cognitive science 14, 2 (1990), 179–211

    1990

Show all 45 references
  1. [9]

    Aritra Ghosh, Neil Heffernan, and Andrew S Lan. 2020. Context-aware atten- tive knowledge tracing. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining. 2330–2339

    2020

  2. [10]

    Bert F. Green. [n. d.]. A general solution for the latent class model of latent structure analysis. Psychometrika 16, 2 ([n. d.]), 151–166. https://doi.org/10. 1007/bf02289112

  3. [11]

    S Hochreiter. 1997. Long Short-term Memory. Neural Computation MIT-Press (1997)

    1997

  4. [12]

    HuEdward J., Yulong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. 2021. LoRA: Low-Rank Adaptation of Large Language Models. arXiv: Computation and Language,arXiv: Computation and Language (Jun 2021)

    2021

  5. [13]

    Unggi Lee, Sungjun Yoon, Joon Seo Yun, Kyoungsoo Park, Younghoon Jung, Damji Stratton, and Hyeoncheol Kim. 2024. Difficulty-Focused Contrastive Learning for Knowledge Tracing with a Large Language Model-Based Difficulty Prediction. In Proceedings of the 2024 Joint Internationa...

    2024

  6. [14]

    Zhaoxing Li, Mark Jacobsen, Lei Shi, Yunzhan Zhou, and Jindi Wang. 2023. Broader and Deeper: A Multi-Features with Latent Relations BERT Knowledge Tracing Model. In Responsive and Sustainable Educational Futures, Olga Viberg, Ioana Jivet, Pedro J. Muñoz-Merino, Maria Perifanou...

    2023

  7. [15]

    Zhaoxing Li, Jujie Yang, Jindi Wang, Lei Shi, and Sebastian Stein. 2024. Integrat- ing lstm and bert for long-sequence data analysis in intelligent tutoring systems. arXiv preprint arXiv:2405.05136 (2024)

    2024 arXiv

  8. [16]

    Qi Liu, Zhenya Huang, Yu Yin, Enhong Chen, Hui Xiong, Yu Su, and Guoping Hu

  9. [17]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019)

    2019 arXiv

  10. [18]

    Yunfei Liu, Yang Yang, Xianyu Chen, Jian Shen, Haifeng Zhang, and Yong Yu

  11. [19]

    Koki Nagatani, Qian Zhang, Masahiro Sato, Yan-Ying Chen, Francine Chen, and Tomoko Ohkuma. 2019. Augmenting knowledge tracing by considering forgetting behavior. In The world wide web conference. 3101–3107

    2019

  12. [20]

    Hiromi Nakagawa, Yusuke Iwasawa, and Yutaka Matsuo. 2019. Graph-based Knowledge Tracing: Modeling Student Proficiency Using Graph Neural Network. In IEEE/WIC/ACM International Conference on Web Intelligence. https://doi.org/ 10.1145/3350546.3352513

    2019

  13. [21]

    Shalini Pandey and George Karypis. 2019. A self-attentive model for knowledge tracing. arXiv preprint arXiv:1907.06837 (2019)

    2019 arXiv

  14. [22]

    Shalini Pandey and Jaideep Srivastava. 2020. RKT: Relation-Aware Self-Attention for Knowledge Tracing. Proceedings of the 29th ACM International Conference on Information & Knowledge Management (2020). https://api.semanticscholar. org/CorpusID:221370579

    2020

  15. [23]

    Chris Piech, Jonathan Bassen, Jonathan Huang, Surya Ganguli, Mehran Sahami, Leonidas J Guibas, and Jascha Sohl-Dickstein. 2015. Deep knowledge tracing. Advances in neural information processing systems 28 (2015)

    2015

  16. [24]

    Shuanghong Shen, Qi Liu, Enhong Chen, Zhenya Huang, Wei Huang, Yu Yin, Yu Su, and Shijin Wang. 2021. Learning process-consistent knowledge tracing. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining. 1452–1460

    2021

  17. [25]

    Shuanghong Shen, Qi Liu, Zhenya Huang, Yonghe Zheng, Minghao Yin, Minjuan Wang, and Enhong Chen. 2024. A survey of knowledge tracing: Models, variants, and applications. IEEE Transactions on Learning Technologies(2024)

    2024

  18. [26]

    Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. 2020. Mpnet: Masked and permuted pre-training for language understanding. Advances in neural information processing systems 33 (2020), 16857–16867

    2020

  19. [27]

    Xiangyu Song, Jianxin Li, Qi Lei, Wei Zhao, Yunliang Chen, and Ajmal Mian. 2022. Bi-CLKT: Bi-graph contrastive learning based knowledge tracing. Knowledge-Based Systems 241 (2022), 108274

    2022

  20. [28]

    Yu Su, Qingwen Liu, Qi Liu, Zhenya Huang, Yu Yin, Enhong Chen, Chris Ding, Si Wei, and Guoping Hu. 2018. Exercise-enhanced sequential modeling for student performance prediction. In Proceedings of the AAAI conference on artificial intelligence, V ol. 32

    2018

  21. [29]

    Jianwen Sun, Mengqi Wei, Jintian Feng, Fenghua Yu, Qing Li, and Rui Zou

  22. [30]

    Weicong Tan, Yuan Jin, Ming Liu, and He Zhang. 2022. BiDKT: Deep Knowledge Tracing with BERT. InAd Hoc Networks and Tools for IT, Wei Bao, Xingliang Yuan, Longxiang Gao, Tom H. Luan, and David Bong Jun Choi (Eds.). Springer International Publishing, Cham, 260–278

    2022

  23. [31]

    Zejie Tiana, Guangcong Zhengc, Brendan Flanaganb, Jiazhi Mic, and Hiroaki Ogatab. 2021. BEKT: deep knowledge tracing with bidirectional encoder repre- sentations from transformers. In Proceedings of the 29th International Conference on Computers in Education, V ol. 2. 6–2

    2021

  24. [32]

    Hanshuang Tong, Yun Zhou, and Zhen Wang. 2020. Exercise hierarchical feature enhanced knowledge tracing. In Artificial Intelligence in Education: 21st Interna- tional Conference, AIED 2020, Ifrane, Morocco, July 6–10, 2020, Proceedings, Part II 21. Springer, 324–328

    2020

  25. [33]

    Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Daniel M

    Hugo Touvron, Louis Martin, Kevin R. Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Daniel M. Bikel, Lukas Blecher, Cristian Cantón Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jer...

    2023 arXiv

  26. [34]

    Ashish Vaswani. 2017. Attention is all you need.arXiv preprint arXiv:1706.03762 (2017)

    2017 arXiv

  27. [35]

    Chenyang Wang, Weizhi Ma, Min Zhang, Chuancheng Lv, Fengyuan Wan, Huijie Lin, Taoran Tang, Yiqun Liu, and Shaoping Ma. 2021. Temporal cross-effects in knowledge tracing. In Proceedings of the 14th ACM International Conference on LLM-KT: Aligning Large Language Models with Know...

    2021

  28. [36]

    Zichao Wang, Angus Lamb, Evgeny Saveliev, Pashmina Cameron, Yordan Za- ykov, José Miguel Hernández-Lobato, Richard E Turner, Richard G Baraniuk, Craig Barton, Simon Peyton Jones, Simon Woodhead, and Cheng Zhang. 2020. Diagnostic questions: The neurips 2020 education challenge....

    2020 arXiv

  29. [37]

    Bihan Xu, Zhenya Huang, Jiayu Liu, Shuanghong Shen, Qi Liu, Enhong Chen, Jinze Wu, and Shijin Wang. 2023. Learning behavior-oriented knowledge tracing. In Proceedings of the 29th ACM SIGKDD conference on knowledge discovery and data mining. 2789–2800

    2023

  30. [38]

    Andrea Zanellati, Daniele Di Mitri, Maurizio Gabbrielli, and Olivia Levrini. 2024. Hybrid models for knowledge tracing: A systematic literature review. IEEE Transactions on Learning Technologies(2024)

    2024

  31. [39]

    Jiani Zhang, Xingjian Shi, Irwin King, and Dit-Yan Yeung. 2017. Dynamic Key- Value Memory Networks for Knowledge Tracing. In Proceedings of the 26th International Conference on World Wide Web. https://doi.org/10.1145/3038912. 3052580 A Prompt Templates In this section, we prov...

    2017 doi

  32. [43]

    How would this calculation be written? Pic290−0 A:8+(2÷5)=2 B:(8+2)÷5=2 C:8+2÷5=2 D:(8+2÷5)=2 Related knowledge concepts: Basic Arithmetic The student answered this question correctly

  33. [44]

    Which symbol belongs in the box? Pic749−0 A:> B:< C:= D:≥ Related knowledge concepts: Basic Arithmetic The student answered this question correctly

  34. [45]

    Basic Arithmetic

    What is the output of this Function Machine? Pic836−0 A:10p B:7p C:5(p+2) D:5p+2 Related knowledge concepts: Writing Expressions The student answered this question incorrectly The target question is given as follows: Tom and Katie are arguing about the result of this Function ...

  35. [2019]

    IEEE Transactions on Knowledge and Data Engineering 33, 1 (2019), 100–115

    Ekt: Exercise-aware knowledge tracing for student performance prediction. IEEE Transactions on Knowledge and Data Engineering 33, 1 (2019), 100–115

    2019

  36. [2021]

    In Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence

    Improving knowledge tracing via pre-training question embeddings. In Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence. 1577–1583

  37. [2024]

    Expert Systems with Applications 238 (2024), 122280

    Progressive knowledge tracing: Modeling learning process from abstract to concrete. Expert Systems with Applications 238 (2024), 122280

    2024

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.