Pith. sign in

REVIEW 4 major objections 5 minor 40 references

A Hierarchical Probabilistic Framework for Incremental Knowledge Tracing in Classroom Settings

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A tree-structured probabilistic model beats transformer and LLM baselines in low-resource, online knowledge tracing.

desk verdict A credible probabilistic model for low-resource KT, but the headline empirical claim currently rests on an unresolved difficulty-label leakage question and missing controlled baselines. read the letter →

arxiv 2506.09393 v1 pith:KSLKFI5M submitted 2025-06-11 cs.CL

Xinyi Gao , Qiucheng Wu , Yang Zhang , Xuechen Liu , Kaizhi Qian , Ying Xu , Shiyu Chang This is my paper · ORCID
classification cs.CL
keywords knowledgetracinghiddenMarkovtreemodelconcepthierarchylow-resourcelearningonlineincrementalstudentmasterymodelingexpectation-maximizationeducationaldatamining
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 usually assumes abundant student history, but real classrooms start cold, stream new answers, and may have few peers. This paper argues that the tree-shaped hierarchy of knowledge concepts—often available in classroom materials—can supply the missing prior. It proposes KT2, a Hidden Markov Tree Model in which each knowledge concept has a hidden mastery variable and each observed answer is an emission from the concept it exercises. Parameters are fit by expectation-maximization on a small shared burn-in set and then updated with a single EM step per new response. On three modules from each of two widely used educational datasets, with 100 students per module and as few as five burn-in exercises per student, KT2 reports the highest AUC, accuracy, and F1 on every module, ahead of online variants of transformer-based models and of 3–7 billion parameter LLMs.

What carries the argument

The load-bearing object is the Hidden Markov Tree Model built on the knowledge-concept tree. Hidden nodes are binary mastery variables per concept, observed variables are binary correctness per exercise, and the transition rule sets that mastering a parent entails mastering all its children, while the emission rule maps mastery to correctness with a difficulty class and a guessing rate. EM with closed-form updates estimates the parameters, the upward-downward algorithm computes the posterior mastery probabilities needed for prediction, and the online scheme runs a single EM iteration per new response after a communal burn-in. This machinery is what lets sparse student histories borrow strength from the tree structure.

What would settle it

Train KT2 on the same burn-in data but with the knowledge-concept tree randomly rewired or with leaf-concept assignments shuffled, and compare AUC on held-out responses; the paper's central claim would be falsified if a permuted tree performs as well as the true tree. A complementary check is to replace the tree with a flat, independent-concepts model and see whether the reported margins over baselines disappear.

Watch

Extended reading notes

Core claim

The paper's central discovery is that a small, interpretable probabilistic model—one that places student mastery on a known tree of knowledge concepts—can outperform much larger neural and LLM competitors exactly in the regime where those competitors are weakest. Concretely, it claims that with 50 target students and as few as five exercises per student in the burn-in set, KT2 achieves the best AUC, accuracy, and F1 across all six simulated classroom modules, and that the advantage persists over burn-in sizes from 5 to 30 exercises. The superiority holds even though the online baseline variants are retrained on the full classroom's cumulative interactions, whereas KT2 updates only from the target student's latest response. Qualitative results show mastery updates propagate along the tree: a correct answer raises posterior mastery not only on the exercised leaf but also on sibling and ancestor concepts.

Load-bearing premise

The load-bearing premise is that the knowledge-concept tree is correct: each exercise involves exactly one concept, each concept has one parent, and knowing a parent concept always means knowing every child concept. If the tree is wrong or these rules do not hold, the model's estimates are conditioned on a mistaken structure.

Editorial extensions

If this is right

  • In a real classroom rollout, KT2 could start giving personalized predictions after roughly five exercises per student plus a shared class-level burn-in.
  • Each new response triggers only one EM update, so online deployment does not require retraining deep models or re-prompting an LLM.
  • Mastery estimates are interpretable per concept: correct answers raise posterior mastery on the exercised leaf and propagate to related concepts, while errors lower them.
  • The reported gains hold across six modules and across burn-in sizes from 5 to 30 exercises per student.

Reading between the lines

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

  • Beyond the paper, this suggests a broader design principle: when domain structure is available, a structured generative prior can substitute for model scale; a direct test would be to compare KT2 with a flat, tree-less Bayesian knowledge-tracing model on the same burn-in data.
  • The paper leaves implicit a natural extension to exercises touching multiple concepts or to non-tree hierarchies; because the authors note their single-leaf, single-parent assumptions, testing graceful degradation under such relaxations would clarify how much the tree structure per se carries the result.
  • Because the MOOCRadar tree is constructed by embedding clustering and LLM labeling rather than expert annotation, an external check would be to rerun KT2 on an expert-built concept tree for the same dataset; if gains shrink, the reported advantage depends on how well the constructed hierarchy captures true cognitive structure.
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 proposes KT2, a probabilistic knowledge tracing model based on a Hidden Markov Tree Model over a hierarchical knowledge-concept (KC) tree. Binary mastery variables at KC nodes follow an entailment-structured transition model, observed responses are generated from mastery via difficulty-dependent emission probabilities, and parameters are estimated by EM with closed-form updates. Inference is performed with an upward-downward algorithm, and an incremental single-EM-step update supports streaming classroom settings. Experiments on six modules from XES3G5M and MOOCRadar compare KT2 with AKT, SAINT, qDKT, and two LLM baselines, reporting consistent AUC, accuracy, and F1 improvements, together with a burn-in-size ablation and qualitative mastery-update visualizations.

Significance. If its empirical claim holds, KT2 is a useful contribution: it shows that a simple, interpretable probabilistic model that exploits tree-structured KC priors can outperform much larger transformer and LLM baselines under cold-start, online, and limited-peer conditions. The closed-form EM updates and the upward-downward inference are clean and should be easy to reproduce, which is a genuine strength. The significance is currently tempered by two evaluation-protocol concerns: the difficulty-label construction may leak target or future information into Eq. (10), and Table 2 contains no error bars or significance tests. Resolving those issues is necessary before the headline claim is fully credible.

major comments (4)
  1. [Sec. 3.1 / Eq. (10)] The paper does not specify how the difficulty labels (easy/medium/hard) are computed. Sec. 3.1 says they are assigned "based on student correctness rates," but it does not say whether those rates come from the full original dataset, from the sampled 100-student classroom, or from a split that excludes each target student's future responses. Since Eq. (10) makes the emission probability phi_n depend on the difficulty class, a label computed from all module interactions would encode the target response and future class responses into the model at inference time. This is a per-question aggregate signal that none of the DLKT-Online or LLM baselines receive, so the reported wins in Table 2 could be an artifact of information leakage rather than of the HMTM. Please state the exact label construction and re-run the comparison with difficulty labels derived only from burn-in data (or from a time-separated split that excludes the predicted response); ideally report both versions.
  2. [Table 2 / Sec. 3.2] All results in Table 2 are point estimates from a single sample of 100 students per module. Many of the reported improvements are small (for example, on the MOOCRadar Wine Knowledge module the AUC gain over qDKT-Online is about 0.005), and the paper provides no standard deviations, confidence intervals, bootstrap resampling, or paired significance tests. Without such measures, the claim that KT2 "consistently outperforms" the baselines is not statistically supported. Please add multiple classroom resamples or bootstrap CIs and appropriate paired tests.
  3. [Abstract / Sec. 3.1] The abstract states that KT2 is evaluated with "50 target students and as few as five exercises per student," but Sec. 3.1 describes sampling 100 students per module and using each student's first 10 interactions as burn-in data. This is a direct factual inconsistency that affects the claimed low-resource regime and the reproducibility of the experiments. Please correct the abstract or the experimental description to match the actual setup.
  4. [Sec. 3.1 / App. E] For MOOCRadar, the knowledge concept tree is constructed by embedding clustering and GPT-4o-mini labeling, with manual annotation for outliers. The paper provides no validation that this automatically constructed hierarchy reflects true prerequisite or entailment structure. Because Eq. (8) assumes that mastering a parent KC entails mastering all its children, a misspecified tree can bias the posterior mastery estimates used in Eq. (15) and may affect the comparison. Please provide some validation (for example, expert agreement on a sample of edges) or a sensitivity analysis over tree-construction choices, or discuss this risk explicitly in the limitations.
minor comments (5)
  1. [Sec. 2.3, Eq. (8)] "not vise versa" is a typo for "not vice versa."
  2. [Table 2] The MOOCRadar dataset heading appears as "MO O CRA D A R" in Table 2; please make the acronym styling consistent (e.g., "MOOCRadar").
  3. [Sec. 3.1] "N( 0.65, 0.15)" should be written as "N(0.65, 0.15)" without the stray space.
  4. [Sec. 2.6] The phrase "performing a single EM iteration" is ambiguous; please clarify whether this means one E-step followed by one M-step, and whether the posterior calculations are run over the full tree for each new observation.
  5. [Appendix A, Eqs. (16)-(18)] The notation in Eq. (18) uses gamma^{(1)}_{ci} and gamma^{(0)}_{ci}, but the definitions in Eqs. (16) and (17) are written with different superscripts; unifying this notation would make the closed-form updates easier to follow.

Circularity Check

1 steps flagged · score 6.0 of 10

The core HMTM derivation is self-contained, but the difficulty-label construction in Eq. (10)/Sec. 3.1 can make the target prediction a function of the target answer itself, undermining the headline empirical claim.

  1. self definitional [Sec. 2.3, Eq. (10); Sec. 3.1 (Data Construction)]
    "Each exercise is annotated with fine-grained KCs and can also be assigned a difficulty label (easy/medium/hard) based on student correctness rates. ... It takes on three possible values,{r_easy,r_med,r_hard}, depending on the difficulty level of the exercise, which is determined by which of the three pre-defined bins (high, medium, or low) the historical solve rate falls into."

    In Eq. (15) the target response Q_{n*} is predicted using phi_{n*} from Eq. (10). Equation (10) sets phi_{n*} from the exercise's difficulty label, which Sec. 3.1 says is assigned based on student correctness rates and which Eq. (10) says is determined by the historical solve rate. The paper never states that this rate is computed only from a pre-split training set or from responses observed before the target/future interaction. Under the natural full-dataset reading, phi_{n*} is a function of the very Q_{n*} being predicted, and of other students' future responses, and the EM estimates of r_easy, r_med, r_hard in Eqs. (22)-(24) are fit on responses partitioned by that same label.

full rationale

The probabilistic derivation itself is internally consistent and not circular: theta is estimated by maximum likelihood on observed responses, and predictions are made by conditioning on student history through Eq. (15). The transition assumption in Eq. (8) is a stated modeling choice rather than a result derived from itself, and the tree construction for MOOCRadar uses external tools and datasets. There are no load-bearing self-citations. The only significant circularity risk is the difficulty-label construction: the paper does not document whether the historical solve rate used to assign each exercise's difficulty is computed from the full dataset, from the sampled classroom, or from data strictly preceding the target response. If the label is precomputed from full-dataset correctness rates, then Eq. (10) makes the emission probability for the target question depend on the target and future correctness, and the reported gains over baselines that receive no such label are not a valid demonstration of the model's predictive derivation. Because this ambiguity affects the paper's central empirical claim, the score is a partial circularity; if the authors confirm that difficulty labels are derived only from burn-in or strictly historical data, the score would drop to 1-2.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a parametric generative model with per-node transition probabilities, three difficulty-dependent emission rates plus a guessing rate, all fit by EM, and on a constructed tree hierarchy. No new physical or conceptual entities are introduced; the model uses standard latent mastery variables.

free parameters (4)
  • gamma_c (per-KC transition probability) = estimated via EM (Eq. 18)
    Probability that a child KC is mastered when the parent is not mastered; one per non-root KC.
  • r_easy, r_med, r_hard (emission probabilities for mastered concepts) = estimated via EM (Eq. 24)
    Correct-answer probabilities for a mastered concept, split by difficulty class.
  • epsilon (guessing probability) = estimated via EM, clipped to <=0.3 (App. G)
    Correct-answer probability when the concept is not mastered; the clip prevents confusion with hard-question emission.
  • Difficulty bin thresholds (easy/medium/hard) = pre-defined but unspecified
    The cutoffs that map historical solve rates to difficulty classes are not stated; they determine which r_l applies to each question in Eq. (10).
assumptions (5)
  • domain assumption Student responses are independent across students, and within a student the hidden mastery states follow a Markov tree with observations conditionally independent given the state (Eqs. 5-9).
    These graphical-model assumptions define the likelihood that is maximized; if dependencies such as peer effects or question-order effects exist, the model is misspecified.
  • domain assumption Mastery of a parent concept entails mastery of all children: p(child=1|parent=1)=1 (Eq. 8).
    This strict entailment semantics is load-bearing for tree propagation; it may not hold for all educational hierarchies.
  • domain assumption Every exercise is associated with exactly one leaf KC and each KC has exactly one parent.
    Stated in Limitations (Sec. 6); many real exercises involve multiple KCs, and DAG structures are common.
  • ad hoc to paper For MOOCRadar, the knowledge tree constructed via embedding clustering, GPT-4o-mini labeling, and manual annotation reflects the true KC hierarchy.
    App. E describes the construction; the resulting tree is a modeling choice specific to this paper, not a given from the dataset.
  • domain assumption Difficulty labels derived from historical solve rates correctly categorize items into easy/medium/hard and are available without using test responses.
    The paper does not specify how difficulty is computed; if computed from the full dataset including target responses, the model's emission prior in Eq. (10) leaks information.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Hierarchical Probabilistic Framework for Incremental Knowledge Tracing in Classroom Settings." pith.science (2026). https://pith.science/paper/KSLKFI5M

@misc{pith2026250609393,
  author       = {Pith},
  title        = {Pith review of: A Hierarchical Probabilistic Framework for Incremental Knowledge Tracing in Classroom Settings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KSLKFI5M}},
  note         = {Machine review of arXiv:2506.09393}
}
abstract

Knowledge tracing (KT) aims to estimate a student's evolving knowledge state and predict their performance on new exercises based on performance history. Many realistic classroom settings for KT are typically low-resource in data and require online updates as students' exercise history grows, which creates significant challenges for existing KT approaches. To restore strong performance under low-resource conditions, we revisit the hierarchical knowledge concept (KC) information, which is typically available in many classroom settings and can provide strong prior when data are sparse. We therefore propose Knowledge-Tree-based Knowledge Tracing (KT$^2$), a probabilistic KT framework that models student understanding over a tree-structured hierarchy of knowledge concepts using a Hidden Markov Tree Model. KT$^2$ estimates student mastery via an EM algorithm and supports personalized prediction through an incremental update mechanism as new responses arrive. Our experiments show that KT$^2$ consistently outperforms strong baselines in realistic online, low-resource settings.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 29 canonical work pages

  1. [1]

    Deep knowledge tracing

    C Piech, J Bassen, J Huang, S Ganguli, M Sahami, LJ Guibas, and J Sohl-Dickstein. Deep knowledge tracing. advances in neural information processing systems. Association for Computing Machinery, pages 201–204, 2015

    work page 2015

  2. [2]

    Context-aware attentive knowledge tracing

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

    work page 2020

  3. [3]

    Dynamic key-value memory networks for knowledge tracing

    Jiani Zhang, Xingjian Shi, Irwin King, and Dit-Yan Yeung. Dynamic key-value memory networks for knowledge tracing. In Proceedings of the 26th international conference on World Wide Web, pages 765–774, 2017

    work page 2017

  4. [4]

    A self-attentive model for knowledge tracing

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

    arXiv 1907

  5. [5]

    simplekt: a simple but tough-to-beat baseline for knowledge tracing

    Zitao Liu, Qiongqiong Liu, Jiahao Chen, Shuyan Huang, and Weiqi Luo. simplekt: a simple but tough-to-beat baseline for knowledge tracing. arXiv preprint arXiv:2302.06881, 2023

    arXiv 2023

  6. [6]

    Explainable Few-shot Knowledge Tracing

    Haoxuan Li, Jifan Yu, Yuanxin Ouyang, Zhuang Liu, Wenge Rong, Juanzi Li, and Zhang Xiong. Explainable few-shot knowledge tracing. arXiv preprint arXiv:2405.14391, 2024

    work page Pith review arXiv 2024

  7. [7]

    Towards modeling learner performance with large language models

    Seyed Parsa Neshaei, Richard Lee Davis, Adam Hazimeh, Bojan Lazarevski, Pierre Dillenbourg, and Tanja Käser. Towards modeling learner performance with large language models. arXiv preprint arXiv:2403.14661, 2024

    arXiv 2024

  8. [8]

    Wavelet-based statistical signal processing using hidden markov models

    Matthew S Crouse, Robert D Nowak, and Richard G Baraniuk. Wavelet-based statistical signal processing using hidden markov models. IEEE Transactions on signal processing, 46(4):886–902, 2002

    work page 2002

Show all 40 references
  1. [9]

    Maximum likelihood from incomplete data via the em algorithm

    Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological), 39(1):1–22, 1977

    1977

  2. [10]

    Xes3g5m: A knowledge tracing benchmark dataset with auxiliary information

    Zitao Liu, Qiongqiong Liu, Teng Guo, Jiahao Chen, Shuyan Huang, Xiangyu Zhao, Jiliang Tang, Weiqi Luo, and Jian Weng. Xes3g5m: A knowledge tracing benchmark dataset with auxiliary information. Advances in Neural Information Processing Systems, 36:32958–32970, 2023

    2023

  3. [11]

    Moocradar: A fine-grained and multi-aspect knowledge repository for improving cognitive student modeling in moocs

    Jifan Yu, Mengying Lu, Qingyang Zhong, Zijun Yao, Shangqing Tu, Zhengshan Liao, Xiaoya Li, Manli Li, Lei Hou, Hai-Tao Zheng, et al. Moocradar: A fine-grained and multi-aspect knowledge repository for improving cognitive student modeling in moocs. In Proceedings of the 46th Int...

    2023

  4. [12]

    Automated knowledge concept anno- tation and question representation learning for knowledge tracing

    Yilmazcan Ozyurt, Stefan Feuerriegel, and Mrinmaya Sachan. Automated knowledge concept anno- tation and question representation learning for knowledge tracing. arXiv preprint arXiv:2410.01727, 2024

    2024 arXiv

  5. [13]

    Umap: Uniform manifold approximation and projection for dimension reduction

    Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018

    2018 arXiv

  6. [14]

    Density-based clustering based on hierarchical density estimates

    Ricardo JGB Campello, Davoud Moulavi, and Jörg Sander. Density-based clustering based on hierarchical density estimates. In Pacific-Asia conference on knowledge discovery and data mining, pages 160–172. Springer, 2013

    2013

  7. [15]

    Gpt-4o mini: advancing cost-efficient intelligence

    OpenAI. Gpt-4o mini: advancing cost-efficient intelligence. https://openai.com/index/gpt-4o- mini-advancing-cost-efficient-intelligence/, 2024

    2024

  8. [16]

    Towards an appropriate query, key, and value computation for knowledge tracing

    Youngduck Choi, Youngnam Lee, Junghyun Cho, Jineon Baek, Byungsoo Kim, Yeongmin Cha, Dongmin Shin, Chan Bae, and Jaewe Heo. Towards an appropriate query, key, and value computation for knowledge tracing. In Proceedings of the seventh ACM conference on learning@ scale, pages 34...

    2020

  9. [17]

    qdkt: Question-centric deep knowledge tracing

    Shashank Sonkar, Andrew E Waters, Andrew S Lan, Phillip J Grimaldi, and Richard G Baraniuk. qdkt: Question-centric deep knowledge tracing. arXiv preprint arXiv:2005.12442, 2020

    2005 arXiv

  10. [18]

    Qwen-Team. Qwen2. 5 technical report. arXiv preprint arXiv:2502.13923, 2025

    2025 arXiv

  11. [19]

    Llama 3.2: Revolutionizing edge ai and vision with open, customizable models

    Meta. Llama 3.2: Revolutionizing edge ai and vision with open, customizable models. https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/, 2024

    2024

  12. [20]

    pykt: a python library to benchmark deep learning based knowledge tracing models

    Zitao Liu, Qiongqiong Liu, Jiahao Chen, Shuyan Huang, Jiliang Tang, and Weiqi Luo. pykt: a python library to benchmark deep learning based knowledge tracing models. Advances in Neural Information Processing Systems, 35:18542–18555, 2022

    2022

  13. [21]

    Item response theory

    Li Cai, Kilchan Choi, Mark Hansen, and Lauren Harrell. Item response theory. Annual Review of Statistics and Its Application, 3(1):297–321, 2016

    2016

  14. [22]

    Deep-irt: Make deep learning based knowledge tracing explainable using item response theory

    Chun-Kit Yeung. Deep-irt: Make deep learning based knowledge tracing explainable using item response theory. arXiv preprint arXiv:1904.11738, 2019

    1904 arXiv

  15. [23]

    Gikt: a graph-based interaction model for knowledge tracing

    Yang Yang, Jian Shen, Yanru Qu, Yunfei Liu, Kerong Wang, Yaoming Zhu, Weinan Zhang, and Yong Yu. Gikt: a graph-based interaction model for knowledge tracing. In Machine learning and knowledge discovery in databases: European conference, ECML PKDD 2020, Ghent, Belgium, Septembe...

    2020

  16. [24]

    Enhancing deep knowledge tracing with auxiliary tasks

    Zitao Liu, Qiongqiong Liu, Jiahao Chen, Shuyan Huang, Boyu Gao, Weiqi Luo, and Jian Weng. Enhancing deep knowledge tracing with auxiliary tasks. In Proceedings of the ACM web conference 2023, pages 4178–4187, 2023

    2023

  17. [25]

    Graph-based knowledge tracing: model- ing student proficiency using graph neural network

    Hiromi Nakagawa, Yusuke Iwasawa, and Yutaka Matsuo. Graph-based knowledge tracing: model- ing student proficiency using graph neural network. In IEEE/WIC/aCM international conference on web intelligence, pages 156–163, 2019

    2019

  18. [26]

    Skill-oriented hierarchical structure for deep knowledge tracing

    Zhenyuan Yang, Shimeng Xu, Changbo Wang, and Gaoqi He. Skill-oriented hierarchical structure for deep knowledge tracing. In 2022 IEEE 34th International Conference on Tools with Artificial Intelligence (ICTAI), pages 425–432. IEEE, 2022

    2022

  19. [27]

    Knowledge structure-aware graph-attention networks for knowledge tracing

    Shun Mao, Jieyu Zhan, Jiawei Li, and Yuncheng Jiang. Knowledge structure-aware graph-attention networks for knowledge tracing. In International Conference on Knowledge Science, Engineering and Management, pages 309–321. Springer, 2022

    2022

  20. [28]

    Deep hierarchical knowledge tracing

    Tianqi Wang, Fenglong Ma, and Jing Gao. Deep hierarchical knowledge tracing. In Proceedings of the 12th international conference on educational data mining, 2019

    2019

  21. [29]

    Hhskt: A learner–question interactions based heterogeneous graph neural network model for knowledge tracing.Expert Systems with Applications, 215:119334, 2023

    Qin Ni, Tingjiang Wei, Jiabao Zhao, Liang He, and Chanjin Zheng. Hhskt: A learner–question interactions based heterogeneous graph neural network model for knowledge tracing.Expert Systems with Applications, 215:119334, 2023

    2023

  22. [30]

    Predictive, scalable and interpretable knowledge tracing on structured domains

    Hanqi Zhou, Robert Bamler, Charley M Wu, and Álvaro Tejero-Cantero. Predictive, scalable and interpretable knowledge tracing on structured domains. In The Twelfth International Conference on Learning Representations, 2024

    2024

  23. [31]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    2023 arXiv

  24. [32]

    Claude 3.7 sonnet and claude code

    Anthropic. Claude 3.7 sonnet and claude code. https://www.anthropic.com/news/claude-3-7- sonnet, 2025

    2025

  25. [33]

    Gemini: Google deepmind’s multimodal llms

    Google DeepMind. Gemini: Google deepmind’s multimodal llms. https://deepmind.google/technologies/gemini/, 2025. 12 Knowledge-Tree-based Knowledge Tracing Preprint

    2025

  26. [34]

    Leveraging large language models for next-generation educational technologies

    Neil Heffernan, Rose Wang, Christopher MacLellan, Arto Hellas, Chenglu Li, Candace Walkington, Joshua Littenberg-Tobias, David Joyner, Steven Moore, Adish Singla, et al. Leveraging large language models for next-generation educational technologies. In Proceedings of the 17th I...

    2024

  27. [35]

    Scaffolding language learning via multi-modal tutoring systems with pedagogical instructions

    Zhengyuan Liu, Stella Xin Yin, Carolyn Lee, and Nancy F Chen. Scaffolding language learning via multi-modal tutoring systems with pedagogical instructions. In 2024 IEEE Conference on Artificial Intelligence (CAI), pages 1258–1265. IEEE, 2024

    2024

  28. [36]

    Chain-of-exemplar: enhancing distractor generation for multimodal educational question generation

    Haohao Luo, Yang Deng, Ying Shen, See-Kiong Ng, and Tat-Seng Chua. Chain-of-exemplar: enhancing distractor generation for multimodal educational question generation. ACL, 2024

    2024

  29. [37]

    Knowledge tracing: Modeling the acquisition of procedural knowledge

    Albert T Corbett and John R Anderson. Knowledge tracing: Modeling the acquisition of procedural knowledge. User modeling and user-adapted interaction, 4:253–278, 1994

    1994

  30. [38]

    Language model can do knowledge tracing: Simple but effective method to integrate language model and knowledge tracing task

    Unggi Lee, Jiyeong Bae, Dohee Kim, Sookbun Lee, Jaekwon Park, Taekyung Ahn, Gunho Lee, Damji Stratton, and Hyeoncheol Kim. Language model can do knowledge tracing: Simple but effective method to integrate language model and knowledge tracing task. arXiv preprint arXiv:2406.02893, 2024

    2024 arXiv

  31. [39]

    C-pack: Packaged resources to advance general chinese embedding, 2023

    Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. C-pack: Packaged resources to advance general chinese embedding, 2023

    2023

  32. [40]

    unsummarizable

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems...

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.