Pith. sign in

REVIEW 3 major objections 4 minor 107 references

Personalized Exercise Recommendation with Semantically-Grounded Knowledge Tracing

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A calibrated knowledge-tracing model can serve as a simulated student, and a model-based critic initialized from it consistently improves RL exercise recommenders in all tasks the authors test.

desk verdict The framework is real and the components are useful, but the headline result is an artifact of scoring the policy on the same KT predictions that generate its rewards. read the letter →

arxiv 2507.11060 v1 pith:TK7ZVW3W submitted 2025-07-15 cs.AI

Yilmazcan Ozyurt , Tunaberk Almaci , Stefan Feuerriegel , Mrinmaya Sachan This is my paper
classification cs.AI
keywords exerciserecommendationknowledgetracingreinforcementlearningmodel-basedvalueestimationcontrastiveconceptannotationlargelanguagemodelspersonalized
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

ExRec claims that personalized exercise recommendation improves when the recommendation policy is trained against a knowledge-tracing (KT) model that understands question content rather than question IDs. The paper builds an end-to-end pipeline: LLMs annotate each question's solution steps and knowledge concepts, contrastive learning embeds questions and concepts in a shared semantic space, and a KT model is calibrated to report a student's mastery of any concept directly. That calibrated model doubles as the RL environment, providing the state, the transition probabilities, and the reward, and a model-based value estimation (MVE) step initializes the critic with the student model's own components. In experiments across four educational tasks on 2,048 students from an online math dataset, MVE consistently improves continuous value-based RL methods, and the framework generalizes to unseen questions. If the KT simulator faithfully tracks real learning, these policies are ready to transfer to real platforms; the main open question is the simulator's fidelity.

What carries the argument

The load-bearing object is the calibrated knowledge-tracing (KT) model acting as the reinforcement-learning environment. Its recurrent state encoder $E_\theta$ compresses the exercise history into a compact latent student state $s_t = E_\theta(s_{t-1}, \tilde{z}^q_t, z^y_t)$, and its predictor $F_\theta$ outputs both next-question correctness and, after KC calibration, direct mastery estimates $\hat{y}^c_t$ for any knowledge concept from its embedding. This one model supplies the MDP's state, its transition probabilities, and the reward $r_t = \hat{y}^c_{t+1} - \hat{y}^c_t$ (Eqs. 18-19). The second mechanism is model-based value estimation: the critic is initialized with the KT model's own components and computes $Q(s_t, \tilde{z}^q_{t+1}) = \hat{y}^q_{t+1} y^{c(1)}_{t+1} + (1 - \hat{y}^q_{t+1}) y^{c(0)}_{t+1} - \hat{y}^c_t$ (Eq. 23), the expected mastery gain under both response outcomes weighted by the KT model's predicted probability, instead of learning the Q-function from trial and error. This gives Q-learning critics a structured, near-zero-cost value estimate, which the paper credits for the consistent gains.

What would settle it

Run a field experiment on a real online math platform: deploy an MVE-trained policy (e.g., DDPG w/ MVE) and a random-policy baseline, and give students held-out probe questions for each targeted knowledge concept before and after the 10-exercise sequence. Compare the measured per-concept mastery change with the KT model's predicted knowledge-state change that the paper uses as its score (Eq. 24). The central claim fails if measured gains do not track predicted gains or if the rank order of policies reverses; because the paper's score is itself computed from the KT model's predictions, this external comparison is necessary.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that KT-guided RL becomes both practical and effective for exercise recommendation under three design choices: questions and knowledge concepts are represented by learned semantic embeddings instead of IDs; the KT model is calibrated to emit concept-level mastery predictions directly, so states and rewards are computed in constant time rather than by scoring the whole question corpus; and Q-learning critics are initialized with the KT model's own transition and prediction components. With these choices, continuous value-based RL methods (DDPG, SAC, TD3) outperform policy-based and discrete-action baselines, and the MVE variant improves each of them on all four tasks, most sharply when the target concept changes every step (weakest-KC remediation). The authors further show that policies trained in this way keep working when the question corpus is tripled with generated questions, and that the resulting per-concept trajectories are interpretable, revealing which knowledge concepts the policy targeted and when. The result is a claim that a KT model, properly calibrated, is a usable simulated student for optimizing and auditing personalized exercise sequencing at scale.

Load-bearing premise

The load-bearing premise is that the KT model's predicted correctness probabilities and predicted knowledge-state changes describe real student learning; the RL policies are trained entirely against this simulated student, so if those predictions are miscalibrated, the recommended exercises could maximize simulated gains while producing little real learning.

Editorial extensions

If this is right

  • Exercise recommenders can run in real time: rewards no longer require scoring every question in the corpus, because the calibrated KT model reads off any concept's mastery directly from its embedding.
  • MVE-trained continuous value-based agents achieve their largest gains when the target concept shifts dynamically, the setting closest to adaptive remediation in live tutoring.
  • New questions can be added to the corpus without retraining the policy: semantically embedded actions generalize, and agents with MVE improved rather than degraded on the tripled corpus.
  • The pipeline removes the manual labeling bottleneck, since LLM annotation aligned with Common Core standards yields consistent concept labels across 7,652 questions at modest API cost.
  • Per-concept learning trajectories make recommendation policies inspectable, letting educators see which knowledge concepts a policy targeted and how predicted mastery moved.

Reading between the lines

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

  • The paper's evaluation is closed-loop: both the reward signal and the reported score come from the KT model's own predictions (Eqs. 19 and 24), so the headline gains are internal to the simulator; a live-platform deployment with external post-test probes is the natural next step.
  • Because the MVE critic is effectively a one-step rollout through the KT model, extending it to multi-step rollouts inside the critic could improve value estimates for recommendation horizons longer than the 10-step episodes tested.
  • The continuous action space plus nearest-question retrieval leaves a discretization gap; a generative model that emits new question embeddings directly could close the loop between recommendation and automated content creation, a direction the extended-corpus experiment gestures toward.
  • A cheap transfer benchmark would compare the KT model's predicted correctness on new questions against real learners' performance on the same questions, isolating where annotation quality or embedding similarity limits prediction.
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 / 4 minor

Summary. The manuscript introduces ExRec, an end-to-end framework for personalized exercise recommendation. It (1) annotates questions with solution steps and knowledge concepts via GPT-4o, (2) learns contrastive question/KC embeddings, (3) trains an LSTM-based knowledge tracing (KT) model with a KC-calibration loss, and (4) formulates exercise recommendation as reinforcement learning (RL) with the KT model as simulator, including a model-based value estimation (MVE) critic initialized from KT components. Experiments on XES3G5M report that continuous value-based RL methods with MVE outperform defaults and discrete methods across four simulated tasks, and that MVE transfers to an extended question corpus. The paper also releases code and annotations.

Significance. The strongest contribution is organizational: a complete modular pipeline with semantic grounding and open-source artifacts. If the KT environment were a validated proxy for real student learning, the MVE result would be a useful demonstration that leveraging a known transition model in the critic speeds RL optimization in recommendation. The KC-level calibration is computationally attractive. However, the significance is conditional: the main evaluation is closed-loop, and the paper's central effectiveness claim for MVE is not yet supported by evidence that the optimized quantity corresponds to real learning. Upstream KT prediction AUC (~81.65) and KC-level MAE (0.028) are reported, but these do not validate recommendation-level outcomes.

major comments (3)
  1. [Section 4.4.1 and Appendix D, Eqs. (19), (23), (24)] The reward and the evaluation metric are both defined as the change in the KT model's predicted knowledge state, and the MVE critic in Eq. (23) is exactly the expected value of that one-step reward under the same model's transition probabilities (Eq. (18)). Consequently, an MVE policy that maximizes Eq. (19) is, by construction, optimizing the quantity in Eq. (24). The comparison against non-MVE RL baselines therefore measures how well each algorithm recovers the reward model, not whether students genuinely learn more. This is load-bearing for the Section 6 claim that MVE "proves especially effective." I request an external validation: for example, measuring improvement on held-out real responses following recommended exercises, or a real-user study; at minimum, a demonstration that the KT-predicted knowledge states track genuine learning outcomes.
  2. [Section 4.3, Eq. (14) and Appendix B] The calibration target y^c_t is the KT model's own average predicted correctness over 20 sampled questions; there is no ground-truth KC mastery label. Thus the "calibrated" knowledge states are internally consistent but not validated against an independent measure of student knowledge. Since knowledge-state gains are the only reported outcome, this circularity affects all four evaluation tasks. Please validate the knowledge-state estimates against an independent measure (for example, later real performance on questions from the same KC) or temper the claims about student learning accordingly.
  3. [Section 6, Figure 3] No error bars, seed variance, or significance tests are reported; with 2048 student environments the means may still be noisy if environment initializations are correlated, so the relative ordering of methods may reflect variance or tuning rather than a real effect. In addition, no comparison is made against existing recommendation methods such as RCL4ER [69], TGKT-RL [14], or concept-aware DKT [4], so the claim that ExRec improves upon prior exercise recommenders is not directly tested. Please add multiple seeds, statistical comparisons, and at least one prior-method baseline.
minor comments (4)
  1. [Section 4.3 and 4.4] Equation (10) defines an averaged question embedding for the exercise history, but the RL action space in Section 4.4 uses the notation \tilde z^q_{t+1} without stating explicitly whether the action is the same averaged embedding or the raw question embedding; please clarify.
  2. [Appendix D] The four tasks are described as "real-world tasks," but they are simulated objectives inside the KT environment; this wording should be adjusted to avoid overstating external validity.
  3. [References] Reference [50] contains a typo ("V olodymyr") that should be corrected.
  4. [Section 2] The sentence "methods such as TGKT-RL [14] require a predefined question-KC graph" has a subject-verb agreement error and should read "requires."

Circularity Check

3 steps flagged · score 8.0 of 10

Reward (Eq. 19), MVE critic (Eq. 23) and evaluation score (Eq. 24) are all the same calibrated-KT prediction, and the KC-calibration target (Eq. 14) is the KT model's own averaged output; reported knowledge gains are self-referential.

  1. self definitional [Sec. 4.4 (Eq. 19) and Sec. D (Eq. 24)]
    "we define the reward as the change in predicted knowledge state, i.e., r_t = y^c_{t+1} - y^c_t ... The evaluation score is computed as the net improvement: Score = 1/|T| sum_{c in T} ( y^c_(10) - y^c_(0) )"

    For Tasks 1-3 the target KC set T is fixed at the start of the episode, so the undiscounted sum of the per-step rewards (Eq. 19) telescopes to exactly the evaluation score (Eq. 24), up to the gamma=0.99 discounting. The RL agent is therefore trained to maximize the same quantity on which it is then judged, with both quantities produced by the same calibrated KT predictor F_theta. The reported 'knowledge improvement' in Figure 3 is the reward function itself, so the central claim that MVE improves performance reduces to how well each policy optimizes the paper's own predicted-knowledge objective; it does not measure real student learning.

  2. self definitional [Sec. 4.4.1 (Eqs. 21-23)]
    "The state transition function E_omega is initialized as E_theta, the recurrent knowledge tracing module from our KT model. The value prediction function F_omega is initialized as F_theta ... Q(s_t, z^q_{t+1}) = y^q_{t+1} * y^c_(t+1,1) + (1 - y^q_{t+1}) * y^c_(t+1,0) - y^c_t"

    Equation 23 is the one-step expected value of the Eq. 19 reward computed from the same KT components that define the reward. Because the MVE critic is initialized with the exact transition and prediction functions used to generate the reward, it starts training with the target objective already built in. Its consistent advantage (Section 6: "w/ MVE proves especially effective") is therefore an expected consequence of this privileged initialization, not independent evidence that the recommended exercises improve learning.

1 more flagged steps
  1. fitted input called prediction [Sec. 4.3 (Eq. 14) and Appendix B (Module 3)]
    "we define the knowledge state of a student for a particular KC as ... y^c_t = E_{q~Q(c)}[y^q_t] ~= 1/|Q(c)| sum_{q in Q(c)} y^q_t ... we approximate a student's knowledge state on a given concept c in C by sampling 20 relevant questions associated with c from the question corpus and averaging the predicted correctness scores."

    The calibration target used as supervision in Eq. 16 is constructed from the KT model's own predicted correctness scores averaged over 20 sampled questions, rather than from any external measurement of knowledge. Training F_theta to reproduce these self-generated averages and then reporting a KC-level MAE of 0.028 (Appendix C) only demonstrates that the calibrated predictor can compress its own outputs. The 'calibrated' knowledge state is thus defined by the model instead of being validated against an independent ground truth.

full rationale

The paper has some genuine external grounding upstream: the KT response predictor is trained on 5.5M real interaction labels and reaches AUC 81.65 (Appendix C), and self-citations to [54] are supporting rather than the main issue. However, the recommendation-level evaluation is closed-loop. The calibrated KT model generates the transition (Eq. 18), supplies the reward (Eq. 19), initializes the MVE critic (Eq. 23), and defines the evaluation metric (Eq. 24); for Tasks 1-3 the evaluation score is the cumulative reward by construction. In addition, the KC knowledge-state calibration target (Eq. 14/B) is the KT model's own averaged prediction, so the 'calibration' is self-distillation. Consequently the central claim that ExRec/MVE improves student knowledge is a claim about optimizing the paper's own predicted-knowledge metric, not about externally verified learning outcomes. A score of 8 reflects that the reported gains reduce by definition to the reward function, even though the upstream KT response prediction provides some independent signal.

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

The central claim rests on a simulator built from a KT model. Several constants (temperature, reward scale, discount, episode length) are hand-chosen, and the calibration target is the KT model's own prediction. No new physical or conceptual entities are introduced.

free parameters (5)
  • contrastive temperature tau = 0.1
    Set in Appendix B; controls the sharpness of the contrastive loss and is chosen by hand.
  • reward scale = 1000
    Table 1; scales the KT-derived reward to stabilize RL training, a hand-chosen constant.
  • discount factor gamma = 0.99
    Table 1; standard RL discount, chosen by hand.
  • questions sampled per KC for calibration = 20
    Appendix B; used to estimate the target knowledge state in Eq. 14; this is a design choice affecting calibration quality.
  • RL episode horizon = 10
    Table 1 and Section 5; the number of recommendations per episode, fixed by the authors.
assumptions (5)
  • standard math Bellman optimality equation for Q-learning
    Used in Eq. 20 for the MVE critic design.
  • domain assumption LLM-generated KC annotations are sufficiently accurate and consistent
    Module 1 relies on GPT-4o output; the paper provides qualitative examples and cites prior preference studies, but no large-scale human validation in this paper.
  • domain assumption Contrastive embeddings align questions and KCs so that nearest-neighbor retrieval is meaningful
    Module 2 and Section 6.1 assume the learned embedding space preserves pedagogical similarity, including for unseen LLM-generated questions.
  • domain assumption The KT model's predictions are a valid simulator of student responses
    Eq. 18 samples correctness from the KT model's probability; if the KT model is miscalibrated, the entire RL environment is wrong.
  • ad hoc to paper The calibration target y^c_t (Eq. 14) can be approximated by the KT model's own predictions over 20 questions
    The 'ground truth' for KC-level mastery is computed from a fixed checkpoint of the same KT model, making the calibration objective self-referential.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Personalized Exercise Recommendation with Semantically-Grounded Knowledge Tracing." pith.science (2026). https://pith.science/paper/TK7ZVW3W

@misc{pith2026250711060,
  author       = {Pith},
  title        = {Pith review of: Personalized Exercise Recommendation with Semantically-Grounded Knowledge Tracing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TK7ZVW3W}},
  note         = {Machine review of arXiv:2507.11060}
}
read the original abstract

We introduce ExRec, a general framework for personalized exercise recommendation with semantically-grounded knowledge tracing. Our method builds on the observation that existing exercise recommendation approaches simulate student performance via knowledge tracing (KT) but they often overlook two key aspects: (a) the semantic content of questions and (b) the sequential, structured progression of student learning. To address this, our ExRec presents an end-to-end pipeline, from annotating the KCs of questions and learning their semantic representations to training KT models and optimizing several reinforcement learning (RL) methods. Moreover, we improve standard Q-learning-based continuous RL methods via a tailored model-based value estimation (MVE) approach that directly leverages the components of KT model in estimating cumulative knowledge improvement. We validate the effectiveness of our ExRec using various RL methods across four real-world tasks with different educational goals in online math learning. We further show that ExRec generalizes robustly to new, unseen questions and that it produces interpretable student learning trajectories. Together, our findings highlight the promise of KT-guided RL for effective personalization in education.

Figures

Figures reproduced from arXiv: 2507.11060 by the authors.

Figure 1
Figure 1. Overview of standard exercise recommendation and its limitations. reward computation requires inference over all questions, which limits real-time applicability; and 4 they typically support only a single RL algorithm. To address these challenges, we introduce ExRec, a novel framework for personalized exercise recommendation with semantically-grounded knowledge tracing. ExRec operates with minimal requirements, rely… view at source ↗
Figure 2
Figure 2. Overview of ExRec framework. Numbered gray arrows show how modules interact. inference over question sets, and 4 supporting a broad range of RL algorithms, including both discrete and continuous action spaces. Unlike prior works, we make our entire pipeline open-source to enable researchers to instantly build and test new exercise recommenders within our framework. 3 Preliminaries Knowledge tracing (KT). KT aims to … view at source ↗
Figure 3
Figure 3. Knowledge improvements across four tasks, averaged over 2048 students in the test set. Our framework supports a range of RL algorithms and enables extensive comparison among methods. 6.1 Use Case: Extending the Question Corpus DDPG SAC TD3 DDPG SAC TD3 20 25 30 35 40 45 50 % of Max Improvement -7.4 +0.1 -14.2 +6.9 +4.3 +0.7 Default Original Corpus Extended Corpus w/ MVE [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Results of extending question corpus. To test whether our framework generalizes to new, unseen questions, we extend the original question corpus by generating three times more questions using GPT-4o (see Appendix G for prompt details and examples). Each generation intr…
Figure 5
Figure 5. Figure 5: Knowledge state evolution for a single student across different policies (Task 4). At each step, the weakest KC is targeted, and the corresponding knowledge trajectory is shown. White dashed boxes mark changes after each recommendation. A KC may appear more than once i…
Figure 6
Figure 6. Figure 6: Distribution of number of KCs per question annotated by our framework. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Most representative KCs across all questions in the dataset. The result is shown after [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Effect of representation learning on question embeddings. Each color indicates a different [PITH_FULL_IMAGE:figures/full_fig_p020_8.png]
Figure 9
Figure 9. Figure 9: Knowledge state evolution for a single student across various policies and non-RL baselines [PITH_FULL_IMAGE:figures/full_fig_p030_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

107 extracted references · 73 canonical work pages

  1. [69]

    Contrastive personalized exercise recommendation with reinforcement learning

    Siyu Wu, Jun Wang, and Wei Zhang. Contrastive personalized exercise recommendation with reinforcement learning. IEEE Transactions on Learning Technologies, 2023

    work page 2023

  2. [14]

    TGKT-based personalized learning path recommendation with reinforcement learning

    Zhanxuan Chen, Zhengyang Wu, Yong Tang, and Jinwei Zhou. TGKT-based personalized learning path recommendation with reinforcement learning. In KSEMt, 2023

    2023

  3. [4]

    Concept-aware deep knowledge tracing and exercise recommendation in an online learning system

    Fangzhe Ai, Yishuai Chen, Yuchun Guo, Yongxiang Zhao, Zhenzhu Wang, Guowei Fu, and Guangyan Wang. Concept-aware deep knowledge tracing and exercise recommendation in an online learning system. EDM, 2019

    2019

  4. [1]

    Knowledge tracing with sequential key-value memory networks

    Ghodai Abdelrahman and Qing Wang. Knowledge tracing with sequential key-value memory networks. In SIGIR, 2019

    2019

  5. [2]

    Knowledge tracing: A survey

    Ghodai Abdelrahman, Qing Wang, and Bernardo Nunes. Knowledge tracing: A survey. ACM Computing Surveys, 2023

    2023

  6. [3]

    Covid-19 pandemic and online learning: The challenges and opportunities

    Olasile Babatunde Adedoyin and Emrah Soykan. Covid-19 pandemic and online learning: The challenges and opportunities. Interactive learning environments, 2023

    2023

  7. [5]

    A taxonomy for learning, teaching, and assessing: A revision of Bloom’s taxonomy of educational objectives: complete edition

    Lorin W Anderson and David R Krathwohl. A taxonomy for learning, teaching, and assessing: A revision of Bloom’s taxonomy of educational objectives: complete edition. 2001

    2001

  8. [6]

    A distributional perspective on reinforcement learning

    Marc G Bellemare, Will Dabney, and Rémi Munos. A distributional perspective on reinforcement learning. In ICML, 2017

    2017

Show all 107 references
  1. [7]

    Dynamic programming

    Richard Bellman. Dynamic programming. Science, 1966

    1966

  2. [8]

    Instrumenting courseware and leveraging data with the Open Learning Initiative (OLI)

    Norman Bier, Stephen Moore, and Martin Van Velsen. Instrumenting courseware and leveraging data with the Open Learning Initiative (OLI). In Learning Analytics & Knowledge, 2019

    2019

  3. [9]

    Openai gym

    Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016

    2016 arXiv

  4. [10]

    Learning path recommendation based on knowledge tracing model and reinforcement learning

    Dejun Cai, Yuan Zhang, and Bintao Dai. Learning path recommendation based on knowledge tracing model and reinforcement learning. In ICCC, 2019

    2019

  5. [11]

    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 Advances in Knowledge Discovery and Data Mining, 2013

    2013

  6. [12]

    Learning factors analysis–a general method for cognitive model evaluation and improvement

    Hao Cen, Kenneth Koedinger, and Brian Junker. Learning factors analysis–a general method for cognitive model evaluation and improvement. In International conference on intelligent tutoring systems, 2006

    2006

  7. [13]

    Improving interpretability of deep sequential knowledge tracing models with question-centric cognitive representations

    Jiahao Chen, Zitao Liu, Shuyan Huang, Qiongqiong Liu, and Weiqi Luo. Improving interpretability of deep sequential knowledge tracing models with question-centric cognitive representations. In AAAI, 2023

    2023

  8. [15]

    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 Learning@Scale, 2020

    2020

  9. [16]

    EdNet: A large-scale hierarchical dataset in education

    Youngduck Choi, Youngnam Lee, Dongmin Shin, Junghyun Cho, Seoyon Park, Seewoo Lee, Jineon Baek, Chan Bae, Byungsoo Kim, and Jaewe Heo. EdNet: A large-scale hierarchical dataset in education. In AIED, 2020

    2020

  10. [17]

    Soft actor-critic for discrete action settings

    Petros Christodoulou. Soft actor-critic for discrete action settings. arXiv preprint arXiv:1910.07207, 2019

    1910 arXiv

  11. [18]

    Cognitive task analysis for expert-based instruction in healthcare

    Richard Clark. Cognitive task analysis for expert-based instruction in healthcare. In Handbook of Research on Educational Communications and Technology. Springer, New York, NY , 2014

    2014

  12. [19]

    Modeling student knowledge: Cognitive tutors in high school and college

    Albert Corbett, Megan McLaughlin, and K Christine Scarpinatto. Modeling student knowledge: Cognitive tutors in high school and college. User modeling and user-adapted interaction, 2000

    2000

  13. [20]

    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, 1994

    1994

  14. [21]

    DGEKT: A dual graph ensemble learning method for knowledge tracing

    Chaoran Cui, Yumo Yao, Chunyun Zhang, Hebo Ma, Yuling Ma, Zhaochun Ren, Chen Zhang, and James Ko. DGEKT: A dual graph ensemble learning method for knowledge tracing. Transactions on Information Systems, 2024

    2024

  15. [22]

    Adaptive and personalized exercise generation for online language learning

    Peng Cui and Mrinmaya Sachan. Adaptive and personalized exercise generation for online language learning. In ACL, 2023. 11

    2023

  16. [23]

    BERT: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019

    2019

  17. [24]

    Metacognitive capabilities of LLMs: An exploration in mathematical problem solving

    Aniket Didolkar, Anirudh Goyal, Nan Rosemary Ke, Siyuan Guo, Michal Valko, Timothy Lillicrap, Danilo Rezende, Yoshua Bengio, Michael Mozer, and Sanjeev Arora. Metacognitive capabilities of LLMs: An exploration in mathematical problem solving. ICML, 2024

    2024

  18. [25]

    Model-based value estimation for efficient model-free reinforcement learning

    Vladimir Feinberg, Alvin Wan, Ion Stoica, Michael I Jordan, Joseph E Gonzalez, and Sergey Levine. Model-based value estimation for efficient model-free reinforcement learning. In ICML, 2018

    2018

  19. [26]

    Addressing function approximation error in actor-critic methods

    Scott Fujimoto, Herke Hoof, and David Meger. Addressing function approximation error in actor-critic methods. In ICML, 2018

    2018

  20. [27]

    Context-aware attentive knowledge tracing

    Aritra Ghosh, Neil Heffernan, and Andrew S Lan. Context-aware attentive knowledge tracing. In KDD, 2020

    2020

  21. [28]

    Future trends in the design strategies and technological affordances of e-learning

    Begoña Gros and Francisco J García-Peñalvo. Future trends in the design strategies and technological affordances of e-learning. In Learning, design, and technology: An international compendium of theory, research, practice, and policy. 2023

    2023

  22. [29]

    Enhancing knowledge tracing via adversarial training

    Xiaopeng Guo, Zhijie Huang, Jie Gao, Mingyu Shang, Maojing Shu, and Jun Sun. Enhancing knowledge tracing via adversarial training. In ACM Multimedia, 2021

    2021

  23. [30]

    Soft actor-critic algorithms and applications

    Tuomas Haarnoja, Aurick Zhou, Kristian Hartikainen, George Tucker, Sehoon Ha, Jie Tan, Vikash Kumar, Henry Zhu, Abhishek Gupta, Pieter Abbeel, et al. Soft actor-critic algorithms and applications. In ICML, 2018

    2018

  24. [31]

    The initial knowledge state of college physics students

    Ibrahim Abou Halloun and David Hestenes. The initial knowledge state of college physics students. American journal of Physics, 1985

    1985

  25. [32]

    Rainbow: Combining improvements in deep reinforcement learning

    Matteo Hessel, Joseph Modayil, Hado Van Hasselt, Tom Schaul, Georg Ostrovski, Will Dabney, Dan Horgan, Bilal Piot, Mohammad Azar, and David Silver. Rainbow: Combining improvements in deep reinforcement learning. In AAAI, 2018

    2018

  26. [33]

    Towards robust knowledge tracing models via k-sparse attention

    Shuyan Huang, Zitao Liu, Xiangyu Zhao, Weiqi Luo, and Jian Weng. Towards robust knowledge tracing models via k-sparse attention. In SIGIR, 2023

    2023

  27. [34]

    Exploring multi-objective exercise recommendations in online education systems

    Zhenya Huang, Qi Liu, Chengxiang Zhai, Yu Yin, Enhong Chen, Weibo Gao, and Guoping Hu. Exploring multi-objective exercise recommendations in online education systems. In CIKM, 2019

    2019

  28. [35]

    Forgetting-aware linear bias for attentive knowledge tracing

    Yoonjin Im, Eunseong Choi, Heejin Kook, and Jongwuk Lee. Forgetting-aware linear bias for attentive knowledge tracing. In CIKM, 2023

    2023

  29. [36]

    Dynamic bayesian networks for student modeling

    Tanja Käser, Severin Klingler, Alexander G Schwing, and Markus Gross. Dynamic bayesian networks for student modeling. Transactions on Learning Technologies, 2017

    2017

  30. [37]

    HiTSKT: A hierarchical transformer model for session-aware knowledge tracing

    Fucai Ke, Weiqing Wang, Weicong Tan, Lan Du, Yuan Jin, Yujin Huang, and Hongzhi Yin. HiTSKT: A hierarchical transformer model for session-aware knowledge tracing. Knowledge-Based Systems, 2024

    2024

  31. [38]

    Personalised and adaptive learning: Emerging learning platforms in the era of digital and smart learning

    Deepak Kem. Personalised and adaptive learning: Emerging learning platforms in the era of digital and smart learning. International Journal of Social Science and Human Research, 2022

    2022

  32. [39]

    Learning path construction using reinforcement learning and bloom’s taxonomy

    Seounghun Kim, Woojin Kim, and Hyeoncheol Kim. Learning path construction using reinforcement learning and bloom’s taxonomy. In Intelligent Tutoring Systems, 2021

    2021

  33. [40]

    Knowledge query network for knowledge tracing: How knowledge interacts with skills

    Jinseok Lee and Dit-Yan Yeung. Knowledge query network for knowledge tracing: How knowledge interacts with skills. In Learning Analytics & Knowledge, 2019

    2019

  34. [41]

    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

  35. [42]

    Continuous control with deep reinforcement learning

    Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. In ICLR, 2016

    2016

  36. [43]

    EKT: Exercise-aware knowledge tracing for student performance prediction

    Qi Liu, Zhenya Huang, Yu Yin, Enhong Chen, Hui Xiong, Yu Su, and Guoping Hu. EKT: Exercise-aware knowledge tracing for student performance prediction. Transactions on Knowledge and Data Engineering, 2019. 12

    2019

  37. [44]

    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. In NeurIPS, 2022

    2022

  38. [45]

    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 The Web Conference, 2023

    2023

  39. [46]

    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. In ICLR, 2023

    2023

  40. [47]

    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. In NeurIPS, 2023

    2023

  41. [48]

    Tracing knowledge state with individual cognition and acquisition estimation

    Ting Long, Yunfei Liu, Jian Shen, Weinan Zhang, and Yong Yu. Tracing knowledge state with individual cognition and acquisition estimation. In SIGIR, 2021

    2021

  42. [49]

    Efl teachers’ perceived efficacy of digital platforms in varied online english classes in a higher education

    Fitriatul Masitoh, Ima Fitriyah, and Irma Khairany Mumtaza. Efl teachers’ perceived efficacy of digital platforms in varied online english classes in a higher education. Journal of Languages and Language Teaching, 2024

    2024

  43. [50]

    Human-level control through deep reinforcement learning

    V olodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. Nature, 2015

    2015

  44. [51]

    Automated generation and tagging of knowledge components from multiple-choice questions

    Steven Moore, Robin Schmucker, Tom Mitchell, and John Stamper. Automated generation and tagging of knowledge components from multiple-choice questions. In Learning@ Scale, 2024

    2024

  45. [52]

    Augmenting knowledge tracing by considering forgetting behavior

    Koki Nagatani, Qian Zhang, Masahiro Sato, Yan-Ying Chen, Francine Chen, and Tomoko Ohkuma. Augmenting knowledge tracing by considering forgetting behavior. In WWW, 2019

    2019

  46. [53]

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

    Hiromi Nakagawa, Yusuke Iwasawa, and Yutaka Matsuo. Graph-based knowledge tracing: Modeling student proficiency using graph neural network. In IEEE/WIC/ACM International Conference on Web Intelligence, 2019

    2019

  47. [54]

    Automated knowledge concept annotation and question representation learning for knowledge tracing

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

    2024 arXiv

  48. [55]

    A self-attentive model for knowledge tracing

    Shalini Pandey and George Karypis. A self-attentive model for knowledge tracing. In EDM, 2019

    2019

  49. [56]

    RKT: Relation-aware self-attention for knowledge tracing

    Shalini Pandey and Jaideep Srivastava. RKT: Relation-aware self-attention for knowledge tracing. In CIKM, 2020

    2020

  50. [57]

    Performance factors analysis–a new alternative to knowledge tracing

    Philip I Pavlik, Hao Cen, and Kenneth R Koedinger. Performance factors analysis–a new alternative to knowledge tracing. In Artificial intelligence in education, pages 531–538. Ios Press, 2009

    2009

  51. [58]

    Deep knowledge tracing

    Chris Piech, Jonathan Bassen, Jonathan Huang, Surya Ganguli, Mehran Sahami, Leonidas J Guibas, and Jascha Sohl-Dickstein. Deep knowledge tracing. In NeurIPS, 2015

    2015

  52. [59]

    Sentence-BERT: Sentence embeddings using siamese BERT-networks

    Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using siamese BERT-networks. In EMNLP, 2019

    2019

  53. [60]

    Trust region policy optimization

    John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region policy optimization. In ICML, 2015

    2015

  54. [61]

    Proximal policy optimization algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  55. [62]

    Assessing student’s dynamic knowledge state by exploring the question difficulty effect

    Shuanghong Shen, Zhenya Huang, Qi Liu, Yu Su, Shijin Wang, and Enhong Chen. Assessing student’s dynamic knowledge state by exploring the question difficulty effect. In SIGIR, 2022

    2022

  56. [63]

    A survey of knowledge tracing: Models, variants, and applications

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

    2024

  57. [64]

    Bi-CLKT: Bi-graph contrastive learning based knowledge tracing

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

    2022

  58. [65]

    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. In EDM, 2020. 13

    2020

  59. [66]

    Knowledge tracing machines: Factorization machines for knowledge tracing

    Jill-Jênn Vie and Hisashi Kashima. Knowledge tracing machines: Factorization machines for knowledge tracing. In AAAI, 2019

    2019

  60. [67]

    Instructions and guide for diagnostic questions: The NeurIPS 2020 education challenge

    Zichao Wang, Angus Lamb, Evgeny Saveliev, Pashmina Cameron, Yordan Zaykov, José Miguel Hernández- Lobato, Richard E Turner, Richard G Baraniuk, Craig Barton, Simon Peyton Jones, et al. Instructions and guide for diagnostic questions: The NeurIPS 2020 education challenge. arXiv...

    2020 arXiv

  61. [68]

    Tianshou: A highly modularized deep reinforcement learning library

    Jiayi Weng, Huayu Chen, Dong Yan, Kaichao You, Alexis Duburcq, Minghao Zhang, Yi Su, Hang Su, and Jun Zhu. Tianshou: A highly modularized deep reinforcement learning library. JMLR, 2022

    2022

  62. [70]

    Large language model augmented exercise retrieval for personalized language learning

    Austin Xu, Will Monroe, and Klinton Bicknell. Large language model augmented exercise retrieval for personalized language learning. In Learning Analytics and Knowledge Conference, 2024

    2024

  63. [71]

    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. EDM, 2019

    2019

  64. [72]

    Addressing two problems in deep knowledge tracing via prediction- consistent regularization

    Chun-Kit Yeung and Dit-Yan Yeung. Addressing two problems in deep knowledge tracing via prediction- consistent regularization. In Learning@Scale, 2018

    2018

  65. [73]

    Tracing knowledge instead of patterns: Stable knowledge tracing with diagnostic transformer

    Yu Yin, Le Dai, Zhenya Huang, Shuanghong Shen, Fei Wang, Qi Liu, Enhong Chen, and Xin Li. Tracing knowledge instead of patterns: Stable knowledge tracing with diagnostic transformer. In The Web Conference, 2023

    2023

  66. [74]

    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 WWW, 2017

    2017

  67. [75]

    Dynamic multi-objective sequence-wise recommendation framework via deep reinforcement learning

    Xiankun Zhang, Yuhu Shang, Yimeng Ren, and Kun Liang. Dynamic multi-objective sequence-wise recommendation framework via deep reinforcement learning. Complex & Intelligent Systems, 2023

    2023

  68. [76]

    Solution_Steps

    Hanqi Zhou, Robert Bamler, Charley M Wu, and Álvaro Tejero-Cantero. Predictive, scalable and inter- pretable knowledge tracing on structured domains. In ICLR, 2024. 14 A Dataset The XES3G5M dataset [ 47] is a large-scale benchmark for knowledge tracing, collected from a real-w...

    2024

  69. [77]

    Identify all the relevant KCs required to solve this problem

  70. [78]

    Justify why each KC is relevant, considering the question and solution steps. 23

  71. [79]

    - List of KCs: Provide a list of unique KCs with the help of your reasoning above, i .e

    You have limited space, so please use 100 words maximum. - List of KCs: Provide a list of unique KCs with the help of your reasoning above, i .e. [<KC 1>, ..., <KC M>]. Don’t enumerate the KCs

  72. [80]

    Provide multiple knowledge concepts only when it is actually needed

  73. [81]

    As the step-by- step solution is already provided, use your judgement to infer which knowledge concept(s) might be needed

    Some questions require a figure, which you won’t be provided. As the step-by- step solution is already provided, use your judgement to infer which knowledge concept(s) might be needed

  74. [82]

    Use your judgement based on your input and your ability to infer how the solution would conclude

    For a small set of solutions, their last step(s) might be missing due to limited token size. Use your judgement based on your input and your ability to infer how the solution would conclude

  75. [83]

    Reasoning

    Remember that knowledge concepts should be appropriate for Math curriculum. If annotated step-by-step solution involves advanced techniques, use your judgment for more simplified alternatives. IMPORTANT NOTE: For your task, try to use the Common Core State Standards for Mathem...

  76. [84]

    Each solution step has to be paired

  77. [85]

    Each knowledge concept has to be paired

  78. [86]

    Map a solution step with a knowledge concept only if they are relevant

  79. [87]

    For instance, if there are 4 solution steps, the pair "5-2" is illegal

    Your pairs cannot contain artificial solution steps. For instance, if there are 4 solution steps, the pair "5-2" is illegal

  80. [88]

    1-1, 1-3, 1-5, 2-4, 3-2, 3-5, 4-2, 4-3, 4-5

    Your pairs cannot contain artificial knowledge concepts. For instance, if there are 3 knowledge concepts, the pair "3-5" is illegal. IMPORTANT: For this field, you will output solution step-knowledge concept pairs in a comma-separated manner and in a single line. For example, ...

  81. [89]

    list_KCs

    Don’t enumerate the KCs or steps within the list. --- Example format below --- {0: {"list_KCs": [<KC 1>, ..., <KC M0>], "question": <question_text>, " list_sol_steps":[<step 1>, ..., <step N0>]}, 1: {"list_KCs": [<KC 1>, ..., <KC M1>], "question": <question_text>, " list_sol_s...

  82. [90]

    Calculate the cost of 32 mechanical pencils: 4 yuan each x 32 = 128 yuan

  83. [91]

    Calculate the cost of 24 pencil boxes: 26 yuan each x 24 = 624 yuan

  84. [92]

    Calculate the cost of 65 pens: 45 yuan each x 65 = 2925 yuan

  85. [93]

    Calculate the cost of 9 school bags: 128 yuan each x 9 = 1152 yuan

  86. [94]

    Add all the costs together: 128 + 624 + 2925 + 1152 = 4829 yuan. • KCs: a) Understanding multiplication, b) Solving multiplication and division problems in real-world contexts, c) Adding and subtracting multi-digit whole numbers Artificial Generation 1 • Question: Eddie is buy...

  87. [95]

    Calculate the cost of 50 boxes of markers: 4 yuan each x 50 = 200 yuan

  88. [96]

    Calculate the cost of 40 packs of paper: 12 yuan each x 40 = 480 yuan

  89. [97]

    Calculate the cost of 20 sets of paints: 30 yuan each x 20 = 600 yuan

  90. [98]

    Calculate the cost of 15 rolls of tape: 8 yuan each x 15 = 120 yuan

  91. [99]

    Add all the costs together: 200 + 480 + 600 + 120 = 1400 yuan

  92. [100]

    Convert the total cost to dollars: 1400 yuan / 7 = 200 dollars. • KCs: a) Understanding multiplication, b) Solving multiplication and division problems in real- world contexts, c) Adding and subtracting multi-digit whole numbers, d) Understanding unit conversion Artificial Gen...

  93. [101]

    Calculate the cost of 20 notebooks: 6 yuan each x 20 = 120 yuan

  94. [102]

    Calculate the cost of 30 folders: 10 yuan each x 30 = 300 yuan

  95. [103]

    Calculate the cost of 10 calculators: 60 yuan each x 10 = 600 yuan. 28

  96. [104]

    Calculate the cost of 5 backpacks: 150 yuan each x 5 = 750 yuan

  97. [105]

    Add all the costs together: 120 + 300 + 600 + 750 = 1770 yuan

  98. [106]

    Calculate the discount: 10% of 1770 yuan = 177 yuan

  99. [107]

    Subtract the discount from the total cost: 1770 - 177 = 1593 yuan. • KCs: a) Understanding multiplication, b) Solving multiplication and division problems in real-world contexts, c) Adding and subtracting multi-digit whole numbers, d) Understanding percentages 29 H Visualizati...

Pith tools

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