Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

TrueReason: An Exemplar Personalised Learning System Integrating Reasoning with Foundational Models

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read TrueReason shows that a large language model composing specialised AI micro-skills can run a personalised learning system, with first evidence from a topic-controlled question generator and a reinforcement-learning recommender.

desk verdict A genuinely useful dataset-construction method for topic-controlled question generation, wrapped in a modular architecture description whose RL evaluation is circular and whose QG numbers are too thin to support the central claims. read the letter →

arxiv 2502.10411 v1 pith:YNY7LEXP submitted 2025-01-23 cs.CY cs.AIcs.CLcs.IRcs.MA

classification cs.CYcs.AIcs.CLcs.IRcs.MA
keywords societyofmindspersonalisedlearningeducationalrecommendersystemstopic-controlledquestiongenerationreinforcementlearnermodellingopenresourceslargelanguagemodels
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

This paper asks whether a personalised learning system can be built the way a human teacher works: not as one giant model, but as a coordinator that calls on specialised 'micro-skills' when needed. The authors present TrueReason, a prototype in which a large language model composes these micro-skills to plan learning journeys, recommend resources, and quiz learners. The concrete claims are that a reinforcement-learning recommender can increase a learner's knowledge of chosen topics in simulation, and that a small pre-trained language model can be fine-tuned to generate questions controlled by a target topic. The reported evidence is a larger semantic gap between prescribed and alternative topics for the question generator (Table 1) and cumulative simulated knowledge gain for the recommender (Figure 7). The value, if true, is an architecture for education AI that is modular, transparent where needed, and cheaper to maintain than a single continually trained model.

What carries the argument

The load-bearing object is the 'society of minds' architecture: a chat-facing LLM (the reasoning agent) with function-call access to micro-skills through an API server, backed by a Wikipedia-based domain model, a TrueLearn learner model, and a PEEKC video collection. The two implemented micro-skills are a Deep Deterministic Policy Gradient recommender that maps approximated knowledge states to knowledge-component vectors and then to videos by cosine similarity, and a topic-controlled question generator obtained by fine-tuning T5-Small on MixSQuAD, a dataset of paired contexts/topics/questions with contrastive mixed-context examples. The architecture is what carries the argument: because micro-skills are separate, the authors can test generalisation of each skill, swap or disable skills, and combine them in new ways without retraining a monolithic model.

What would settle it

Run the DDPG policy against real informal learners on the same educational video collection: record TrueLearn's predicted engagement and knowledge gain, then compare them with actual engagement and post-viewing quiz scores on the recommended resources; if simulated gains do not track real outcomes, the recommender's claimed capability is unsupported. For the question generator, have teachers identify which of two topics a generated question targets; if they cannot pick the prescribed topic more often than chance, the WikiSemRel gap is not evidence of educational topical control.

Watch

Extended reading notes

Core claim

The central claim is that a 'society of minds'—an LLM orchestrator composing specialised micro-skills—can operationalise planning and reasoning for personalised learning. On the question-generation side, the paper asserts feasibility of fine-tuning a pre-trained language model (T5-Small) for topic-controlled question generation (T-CQG), using a new contrastive dataset (MixSQuAD) built by wikifying SQuAD contexts and questions. The TopicQG model matches the baseline on BERTScore for target questions while widening the WikiSemRel gap between prescribed and alternative topics from 0.545 to 0.595, which the paper reads as stronger topical control. On the recommendation side, a DDPG agent trained against an environment where TrueLearn simulates engagement and updates a ground-truth knowledge state produces cumulative increases in simulated knowledge over 100 recommended resources. The paper frames these as first steps rather than finished products, stating that the architecture itself is the demonstration.

Load-bearing premise

The load-bearing premise is that TrueLearn's predicted engagement, applied to a simulated ground-truth knowledge state and then used to compute both reward and reported knowledge gain, faithfully stands in for how real informal learners engage with and learn from recommended videos.

Editorial extensions

If this is right

  • If the society-of-minds architecture works, a lifelong learning assistant can move from single-activity recommendations (watch a video) to activity-rich pathways: recommend a video, quiz the learner, review gaps, and explain prerequisites.
  • The T-CQG result implies teachers and learning platforms can generate topic-specific questions from small, fine-tuned models at low cost, reducing dependence on large commercial language models.
  • The MixSQuAD contrastive method implies that topical control can be taught to a small language model without new human annotations, only by recombining existing question datasets through wikification.
  • If the simulated RL recommender reflects real engagement, the system can plan multi-step learning trajectories over prerequisite and interest topics without hand-crafted curricula.
  • The modular micro-skill design implies new skills (summarisation, analogies, feedback) can be added over time without risking unlearning in the rest of the system.

Reading between the lines

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

  • The paper's RL evidence is entirely simulated: knowledge gain is computed inside TrueLearn from a ground-truth state, so the claimed learning benefit would need testing with real informal learners before being taken as an effect on human knowledge.
  • The MixSQuAD contrastive recipe could likely transfer to other constrained generation tasks—summaries, explanations, distractors—where a topic or concept must control generation, since it only needs an existing dataset and an entity linker.
  • A remaining question the paper leaves open is whether Wikipedia-concept relatedness captures pedagogical prerequisites; the domain graph labels relations with a large language model but the system does not yet validate those labels as learning prerequisites.
  • If small language models can match the topical control of larger ones on this task, the same evaluation setup (prescribed versus alternative topic gap) could become a cheap screening test for controlled generation more broadly.
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 / 6 minor

Summary. The paper presents TrueReason, a prototype personalised learning system in which an LLM chat assistant orchestrates specialised 'micro-skills' over a Wikipedia-based domain model, a TrueLearn learner model, and a collection of educational videos. Two micro-skills are described with experiments: (1) a DDPG-based multi-step recommender trained in a simulated environment where two TrueLearn instances act as a human proxy, and (2) a topic-controlled question generation (T-CQG) system built by fine-tuning T5-small on new datasets (SQuAD+, MixSQuAD) derived from SQuAD via Wikification, evaluated on MixKhanQ. The authors claim feasibility of T-CQG and capability of the RL recommender to increase learner knowledge in specific interests, and they propose the modular architecture as a step toward a 'society of minds' AI for education.

Significance. The modular 'society of minds' architecture is a timely and useful framing for sustainable, transparent AI in education; the use of Wikipedia entity linking for a scalable domain model is pragmatic; and the SQuAD+/MixSQuAD dataset construction with small-model fine-tuning for topic-controlled QG is a concrete, reproducible-in-principle contribution, although code and data are not provided. The paper also honestly acknowledges that generated questions lack answerability and that the system is an early prototype. If the architecture and the T-CQG result hold up, the work is a useful proof-of-concept for composing small specialised models rather than relying on a monolithic LLM. However, the experimental support is currently insufficient for the RL recommender claim, and the orchestration claim is not directly evaluated.

major comments (4)
  1. [§4.1.2, Eq. (3), §4.1.4, Fig. 7] The reported 'knowledge gain' in Figure 7 is the same cumulative quantity that DDPG is trained to maximise. Eq. (3) defines the reward as the per-step sum of GTKS knowledge-component changes for the learner's interests, and §4.1.4 states that the reward is calculated from the GTKS update; Figure 7 plots that cumulative increase. The figure therefore shows that the policy optimised its training reward, not that a learner's knowledge increased. The conclusion in §4.1.5 that the recommender 'has been shown to be capable of recommending a sequence of resources to increase a learner's knowledge state' is unsupported by this evidence. Please add a non-optimized baseline (e.g., random or TrueLearn novelty-based recommendation), report held-out evaluation with varying simulated learner profiles, or explicitly reframe the claim as demonstrating optimisation of the TrueLearn proxy reward.
  2. [§4.1.2, §3.6] The simulated environment assumes that TrueLearn's engagement predictions and GTKS update rules are a valid proxy for how informal learners engage and learn, but no evidence for this assumption is provided. The GTKS is initialised by the authors and updated according to TrueLearn's internal rules, so the 'ground-truth' state is itself a model output. Without validation against real learner interaction data, even the PEEKC dataset that the system uses elsewhere, the RL recommender's reported behaviour is a property of the simulator. The paper should either provide such validation or explicitly scope the claim to the simulator rather than to real learner learning.
  3. [§4.2.6, Table 1] The T-CQG evaluation lacks error bars, significance tests, and the number of evaluation examples; Table 1 reports single-point metrics. The BERTScore for target questions is identical (0.859) between Baseline and TopicQG, and the TopicQG advantage is only 0.029 on the discriminating BERTScore gap and 0.050 on the WikiSemRel gap. Since WikiSemRel and the training data construction both rely on Wikipedia entity linking, the improvement may partly reflect alignment with Wikipedia concepts rather than pedagogically meaningful topic control. A human evaluation of topical relevance, answerability, and educational usefulness, or at least a statistical analysis over seeds and evaluation samples, is needed to support the feasibility claim in §4.2.2.
  4. [§3.4, §6] The central architectural claim, that an LLM can compose specialised micro-skills to operationalise planning and reasoning, is not evaluated. The evidence is limited to a scripted dialogue example and the assertion that the assistant 'shows the ability to orchestrate' skills. There is no task-based test of whether the LLM selects the correct micro-skill, respects learner state, or maintains a coherent multi-step plan. Please add an evaluation of the orchestration layer, even a small user study or scenario-based trace, or weaken the central claim to a system proposal.
minor comments (6)
  1. [§4.2.6] The text says the automated evaluation 'addresses RQs 2-5' and that six models including 8-bit and 4-bit versions were developed, but §4.2.2 defines only one research question and §4.2.5 describes only two models. Please reconcile these statements.
  2. [§4.2.3, §4.2.4] The text states that 'three novel datasets' are created but lists only SQuAD+ and MixSQuAD; MixKhanQ is introduced earlier as the evaluation set. Clarify the dataset taxonomy.
  3. [§4.3, Table 2] Table 2 is referenced as evidence that generated questions lack answerability, but the table content is not present in the manuscript. Include the sample generations or remove the reference.
  4. [Various] The manuscript contains several typos and unresolved citation placeholders: 'BERTSCore' for 'BERTScore' (§4.2.6), 'the the' (§4.1), 'Wikpedia' (§3.4), 'monolythic' (§5.3), 'langauge' (§6), 'bahavior' (§5.3), and '[?]' / '[ ?]' in §4.2. A thorough proofread is needed.
  5. [§4.2.5] Please report the validation split, early stopping criterion, number of training steps, and random seeds for the T5 fine-tuning, since only the optimizer hyperparameters are given.
  6. [Figure 7] The curves in Figure 7 presumably average five runs, but no error bars or dispersion measures are shown; indicate whether the plotted values are means and add error bars or interquartile ranges.

Circularity Check

1 steps flagged · score 6.0 of 10

The RL recommender's reported 'knowledge gain' is the cumulative GTKS reward that DDPG was trained to maximize, so Figure 7 is a self-consistency check; T-CQG remains an independent evaluation.

  1. fitted input called prediction [Section 4.1.2 (Eq. 3), Section 4.1.3, Section 4.1.4 (Figure 7), Section 4.1.5]
    "The reward function R(s_t,a_t) can be set to the following in order to take the learners interests into account: R(s_t,a_t) = ∑_{i∈I}(KC^i_t − KC^i_{t′}) ... The GTKS is used strictly for engagement prediction using the TrueLearn model and the reward function is calculated based on the change in the GTKS after updating the GTKS as per TrueLearns rules at each time step ... The results show the cumulative increase in in knowledge for a set of topics of interest after 50 000 training steps."

    Equation 3 defines the reward as the per-step GTKS increase for the learner's interests, and DDPG is trained to maximize this total reward (Section 4.1.3). Figure 7 reports the cumulative GTKS increase as 'knowledge gain,' and Section 4.1.5 concludes the recommender 'has been shown to be capable of recommending a sequence of resources to increase a learners knowledge state in specific interests.' The reported outcome is therefore the optimized objective itself: the policy is fit to maximize exactly the quantity later presented as evidence. Since the GTKS is a TrueLearn simulation and the same state generates both reward and reported gain, this is a self-consistency check, not an external validation of learning.

full rationale

Two micro-skills are presented. The T-CQG chain is independent: topics are obtained by Wikification, TopicQG is fine-tuned on MixSQuAD, and evaluation uses MixKhanQ (a held-out KhanQ-derived set); WikiSemRel is Wikipedia-based but is not the training objective, so Table 1 is a genuine, if narrow, comparison. The RL recommender, however, is circular in evaluation. Eq. 3 sets the reward to the per-step GTKS change; the paper explicitly says the reward is computed from the GTKS after each TrueLearn update. DDPG is trained to maximize this total reward, and Figure 7 reports the cumulative GTKS increase as 'knowledge gain,' with Section 4.1.5 concluding the recommender can increase the learner's knowledge state. The reported capability is the optimized objective itself, and no external human-engagement validation is offered; the authors' own admission that future work must check the qualitative flow of resources confirms the current evidence is a simulation-internal consistency result. Self-citations to TrueLearn and PEEK are prior published models with external data and are not the source of circularity. Because the central RL claim reduces by construction while the T-CQG claim retains independent content, the score is 6.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a simulated learner (TrueLearn as human proxy), an assumed Wikipedia-based notion of topic, and an unevaluated LLM orchestrator. These are domain assumptions rather than fitted parameters; the main unreported tuning choices are the RL hyperparameters.

free parameters (2)
  • Reinforcement learning discount factor gamma = not reported
    Appears in Eq. 4 and the discounted expected return definition, but no value is given in Section 4.1.3; the choice affects the learned recommendation policy.
  • DDPG hyperparameters (learning rate, replay buffer size, exploration noise, target update frequency) = not reported
    Actor and critic architecture is described as three ReLU layers each, but training hyperparameters are omitted, so exact reproduction is not possible.
assumptions (3)
  • domain assumption TrueLearn's engagement prediction is a valid proxy for real learner engagement and knowledge updates
    Used to construct the simulated training environment and to compute the reward and evaluation in Sections 4.1.2 and 4.1.4; the paper states this assumption explicitly.
  • domain assumption Wikipedia concepts and Wikification-based semantic relatedness capture pedagogically meaningful topics
    Topics in SQuAD+/MixSQuAD are chosen by PageRank of Wikified concepts (Section 4.2.4) and the WikiSemRel evaluation metric (Section 4.2.6) uses the same Wikipedia link structure.
  • domain assumption A large language model can reliably orchestrate micro-skills and maintain a coherent pedagogical dialogue
    The prototype uses ChatGPT's Assistant API as the reasoning/planning agent (Sections 3.2 and 3.4), but no evaluation of dialogue quality or planning accuracy is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TrueReason: An Exemplar Personalised Learning System Integrating Reasoning with Foundational Models." pith.science (2026). https://pith.science/paper/YNY7LEXP

@misc{pith2026250210411,
  author       = {Pith},
  title        = {Pith review of: TrueReason: An Exemplar Personalised Learning System Integrating Reasoning with Foundational Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YNY7LEXP}},
  note         = {Machine review of arXiv:2502.10411}
}
read the original abstract

Personalised education is one of the domains that can greatly benefit from the most recent advances in Artificial Intelligence (AI) and Large Language Models (LLM). However, it is also one of the most challenging applications due to the cognitive complexity of teaching effectively while personalising the learning experience to suit independent learners. We hypothesise that one promising approach to excelling in such demanding use cases is using a \emph{society of minds}. In this chapter, we present TrueReason, an exemplar personalised learning system that integrates a multitude of specialised AI models that can mimic micro skills that are composed together by a LLM to operationalise planning and reasoning. The architecture of the initial prototype is presented while describing two micro skills that have been incorporated in the prototype. The proposed system demonstrates the first step in building sophisticated AI systems that can take up very complex cognitive tasks that are demanded by domains such as education.

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. Catching The Correct Answer Trap: Characterising AI Tutor Blind Spots When Analysing Student Reasoning

    cs.CY 2026-04 conditional novelty 6.0 of 10

    AI tutoring models systematically fail to detect student misconceptions when flawed reasoning coincidentally produces the correct answer, with 71% of failures concentrated in two predictable question types.

Reference graph

Works this paper leans on

72 extracted references · 60 canonical work pages · cited by 1 Pith paper

  1. [1]

    ACM Computing Surveys 55(11), 1–37 (2023)

    Abdelrahman, G., Wang, Q., Nunes, B.: Knowledge tracing: A survey. ACM Computing Surveys 55(11), 1–37 (2023)

  2. [2]

    MIS Quarterly42(1), 313–A7 (2018)

    Bauman, K., Tuzhilin, A.: Recommending remedial learning materials to students by filling their knowledge gaps. MIS Quarterly42(1), 313–A7 (2018)

  3. [3]

    Journal of Machine Learning Research 3 (2003)

    Blei, D.M., Ng, A.Y., Jordan, M.I.: Latent dirichlet allocation. Journal of Machine Learning Research 3 (2003)

  4. [4]

    In: Proc

    Blobstein, A., Izmaylov, D., Yifat, T., Levy, M., Segal, A.: Angel: A new generation tool for learning material based questions and answers. In: Proc. of the NeurIPS Workshop on Generative AI for Education (GAIED)

  5. [5]

    In: Proc

    Brank, J., Leban, G., Grobelnik, M.: Annotating documents with relevant wikipedia concepts. In: Proc. of Slovenian KDD Conference on Data Mining and Data Warehouses (SiKDD) (2017)

  6. [6]

    In: AAAI Conference on Artificial Intelligence (2020)

    Bulathwela, S., Perez-Ortiz, M., Yilmaz, E., Shawe-Taylor, J.: Truelearn: A family of bayesian algorithms to match lifelong learners to open educational resources. In: AAAI Conference on Artificial Intelligence (2020)

  7. [7]

    In: Proc

    Bulathwela, S., Kreitmayer, S., P ´erez-Ortiz, M.: What’s in It for Me? Augmenting Recom- mended Learning Resources with Navigable Annotations. In: Proc. of the Int. Conf. on Intel- ligent User Interfaces Companion (2020)

  8. [8]

    In: Proc

    Bulathwela, S., Muse, H., Yilmaz, E.: Scalable educational question generation with pre- trained language models. In: Proc. of Int. Conf. on Artificial Intelligence in Education. pp. 327–339. Springer (2023)

Show all 72 references
  1. [9]

    CoRRabs/2109.03154 (2021), https: //arxiv.org/abs/2109.03154

    Bulathwela, S., P ´erez-Ortiz, M., Novak, E., Yilmaz, E., Shawe-Taylor, J.: PEEK: A large dataset of learner engagement with educational videos. CoRRabs/2109.03154 (2021), https: //arxiv.org/abs/2109.03154

  2. [10]

    CoRR abs/2112.04368 (2021), https://arxiv.org/abs/2112.04368

    Bulathwela, S., P ´erez-Ortiz, M., Yilmaz, E., Shawe-Taylor, J.: Semantic TrueLearn: Using Semantic Knowledge Graphs in Recommendation Systems. CoRR abs/2112.04368 (2021), https://arxiv.org/abs/2112.04368

  3. [11]

    In: Semantic AI in Knowledge Graphs, pp

    Bulathwela, S., P ´erez-Ortiz, M., Yilmaz, E., Shawe-Taylor, J.: Leveraging semantic knowledge graphs in educational recommenders to address the cold-start problem. In: Semantic AI in Knowledge Graphs, pp. 1–20. CRC Press (2023)

  4. [12]

    Sustainability 16(2) (2024)

    Bulathwela, S., P ´erez-Ortiz, M., Holloway, C., Cukurova, M., Shawe-Taylor, J.: Artificial in- telligence alone will not democratise education: On educational inequality, techno-solutionism and inclusive tools. Sustainability 16(2) (2024)

  5. [13]

    In: Mitro- vic, A., Bosch, N

    Bulathwela, S., Verma, M., Ortiz, M.P., Yilmaz, E., Shawe-Taylor, J.: Can population- based engagement improve personalisation? A novel dataset and experiments. In: Mitro- vic, A., Bosch, N. (eds.) Proceedings of the 15th International Conference on Educational Data Mining. pp...

  6. [14]

    Sustainability 14(18) (2022)

    Bulathwela, Sahan and P ´erez-Ortiz, Mar´ıa and Yilmaz, Emine and Shawe-Taylor, John: Power to the Learner: Towards Human-Intuitive and Integrative Recommendations with Open Edu- cational Resources. Sustainability 14(18) (2022)

  7. [15]

    In: The 3rd workshop on meta- cognition and self-regulated learning in educational technologies, at ITS2008

    Bull, S., Kay, J.: Metacognition and open learner models. In: The 3rd workshop on meta- cognition and self-regulated learning in educational technologies, at ITS2008. pp. 7–20 (2008)

  8. [16]

    International Journal of Artificial Intelligence in Education 26(1), 293–331 (2016)

    Bull, S., Kay, J.: Smili: a framework for interfaces to learning data in open learner models, learning analytics and related fields. International Journal of Artificial Intelligence in Education 26(1), 293–331 (2016)

  9. [17]

    In: Cohn, T., He, Y., Liu, Y

    Cachola, I., Lo, K., Cohan, A., Weld, D.: TLDR: Extreme summarization of scientific documents. In: Cohn, T., He, Y., Liu, Y. (eds.) Findings of the Association for Compu- tational Linguistics: EMNLP 2020. pp. 4766–4777. Association for Computational Lin- guistics, Online (Nov ...

  10. [18]

    In: Proc

    Chaplot, D.S., MacLellan, C., Salakhutdinov, R., Koedinger, K.: Learning cognitive models using neural networks. In: Proc. of Artificial Intelligence in Education (2018)

  11. [19]

    User Modeling and User-Adapted Interaction4(4) (1994)

    Corbett, A.T., Anderson, J.R.: Knowledge tracing: Modeling the acquisition of procedural knowledge. User Modeling and User-Adapted Interaction4(4) (1994)

  12. [20]

    In: International Conference on Learning Representations (2020), https://openreview.net/forum?id= H1edEyBKDS

    Dathathri, S., Madotto, A., Lan, J., Hung, J., Frank, E., Molino, P., Yosinski, J., Liu, R.: Plug and play language models: A simple approach to controlled text generation. In: International Conference on Learning Representations (2020), https://openreview.net/forum?id= H1edEyBKDS

  13. [21]

    arXiv preprint arXiv:2402.01580 (2024)

    Denny, P., Gulwani, S., Heffernan, N.T., K ¨aser, T., Moore, S., Rafferty, A.N., Singla, A.: Generative ai for education (gaied): Advances, opportunities, and challenges. arXiv preprint arXiv:2402.01580 (2024)

  14. [22]

    In: Proc

    Du, X., Shao, J., Cardie, C.: Learning to ask: Neural question generation for reading com- prehension. In: Proc. Annual Meeting of the Association for Computational Linguistics. pp. 1342–1352 (2017)

  15. [23]

    In: AAAI Conference on Artificial Intelligence (2024)

    Elkins, S., Kochmar, E., Cheung, J.C., Serban, I.: How teachers can use large language models and bloom’s taxonomy to create educational quizzes. In: AAAI Conference on Artificial Intelligence (2024)

  16. [24]

    In: Artificial Intelligence in Education

    Fawzi, F., Balan, S., Cukurova, M., Yilmaz, E., Bulathwela, S.: Towards human-like educa- tional question generation with small language models. In: Artificial Intelligence in Education. Posters and Late Breaking Results, Workshops and Tutorials, Industry and Innovation Tracks...

  17. [25]

    In: Proc

    Fawzi, F., Amini, S., Bulathwela, S.: Small generative language models for educational question generation. In: Proc. of the NeurIPS Workshop on GAIED

  18. [26]

    In: Proceedings of the 19th ACM International Conference on Information and Knowledge Management

    Ferragina, P., Scaiella, U.: Tagme: on-the-fly annotation of short text fragments (by wikipedia entities). In: Proceedings of the 19th ACM International Conference on Information and Knowledge Management. p. 1625–1628. CIKM ’10, Association for Computing Machinery, New York, N...

  19. [27]

    In: Proceedings of the 29th International Conference on Computational Linguistics (2022)

    Gong, H., Pan, Liangming andHu, H.: KHANQ: A dataset for generating deep questions in education. In: Proceedings of the 29th International Conference on Computational Linguistics (2022)

  20. [28]

    Computers & Education 154, 103878 (2020)

    Hooshyar, D., Pedaste, M., Saks, K., Leijen, ¨A., Bardone, E., Wang, M.: Open learner mod- els in supporting self-regulated learning in higher education: A systematic literature review. Computers & Education 154, 103878 (2020)

  21. [29]

    Manuscrit de Th`ese page 99

    Ilievski, F., Shenoy, K., Klein, N., Chalupsky, H., Szekely, P.: Capturing concept similarity with knowledge graphs. Manuscrit de Th`ese page 99

  22. [30]

    In: The Semantic Web – ISWC 2021

    Ilkou, E., Abu-Rasheed, H., Tavakoli, M., Hakimov, S., Kismih ´ok, G., Auer, S., Nejdl, W.: Educor: An educational and career-oriented recommendation ontology. In: The Semantic Web – ISWC 2021. pp. 546–562. Springer International Publishing, Cham (2021) TrueReason: Integrating...

  23. [31]

    In: Proceedings of the 17th International Conference on Educational Data Mining

    Jia, Q., Cui, J., Du, H., Rashid, P., Xi, R., Li, R., Gehringer, E.: Llm-generated feedback in real classes and beyond: Perspectives from students and instructors. In: Proceedings of the 17th International Conference on Educational Data Mining. pp. 862–867 (2024)

  24. [32]

    Information Systems 65 (2017)

    Kang, J., Lee, H.: Modeling user interest in social media using news media and wikipedia. Information Systems 65 (2017). https://doi.org/https://doi.org/10.1016/j.is.2016.11.003

  25. [33]

    In: International Conference on Learning Representations (2021), https:// openreview.net/forum?id=jWkw45-9AbL

    Khalifa, M., Elsahar, H., Dymetman, M.: A distributional approach to controlled text generation. In: International Conference on Learning Representations (2021), https:// openreview.net/forum?id=jWkw45-9AbL

  26. [34]

    Applied Intelligence 48, 1519–1535 (2018)

    Kla ˇsnja-Mili´cevi´c, A., Ivanovi´c, M., Vesin, B., Budimac, Z.: Enhancing e-learning systems with personalized recommendation based on collaborative tagging techniques. Applied Intelligence 48, 1519–1535 (2018)

  27. [35]

    Future Internet 16(1), 12 (2023)

    Li, X., Henriksson, A., Duneld, M., Nouri, J., Wu, Y.: Evaluating embeddings from pre- trained language models and knowledge graphs for educational content recommendation. Future Internet 16(1), 12 (2023)

  28. [36]

    arXiv preprint arXiv:1509.02971 (2015)

    Lillicrap, T.: Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971 (2015)

  29. [37]

    In: Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N.D., Weinberger, K.Q

    Lindsey, R.V., Khajah, M., Mozer, M.C.: Automatic discovery of cognitive skills to improve the prediction of student learning. In: Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N.D., Weinberger, K.Q. (eds.) Advances in Neural Information Processing Systems 27, pp. 1386–13...

  30. [38]

    In: LREC 2020 - 12th Language Resources and Evaluation Conference

    Martin, L., Villemonte de La Clergerie, ´E., Sagot, B., Bordes, A.: Controllable Sentence Sim- plification. In: LREC 2020 - 12th Language Resources and Evaluation Conference. Marseille, France (May 2020), https://inria.hal.science/hal-02678214

  31. [39]

    In: International Conference on Breaking Barriers with Generative Intelligence

    Miladi, F., Psych ´e, V., Lemire, D.: Comparative performance of gpt-4, rag-augmented gpt-4, and students in moocs. In: International Conference on Breaking Barriers with Generative Intelligence. pp. 81–92. Springer (2024)

  32. [40]

    Modran, H., Bogdan, I.C., Ursut,iu, D., Samoila, C., Modran, P.L.: Llm intelligent agent tutoring in higher education courses using a rag approach (2024)

  33. [41]

    In: Proceedings of the OER20: Open Education Conference (2020)

    Molan, M., Bulathwela, S., Orlic, D.: Accessibility recommendation system. In: Proceedings of the OER20: Open Education Conference (2020)

  34. [42]

    Scientific Reports 11(1), 13943 (2021)

    Moutsinas, G., Shuaib, C., Guo, W., Jarvis, S.: Graph hierarchy: a novel framework to analyse hierarchical structures in complex networks. Scientific Reports 11(1), 13943 (2021)

  35. [43]

    In: Extended Abstracts of the CHI Conference on Human Factors in Computing Systems

    Park, M., Kim, S., Lee, S., Kwon, S., Kim, K.: Empowering personalized learning through a conversation-based tutoring system with student modeling. In: Extended Abstracts of the CHI Conference on Human Factors in Computing Systems. pp. 1–10 (2024)

  36. [44]

    P ´erez Ortiz, M., Bulathwela, S., Dormann, C., Verma, M., Kreitmayer, S., Noss, R., Shawe- Taylor, J., Rogers, Y., Yilmaz, E.: Watch less and uncover more: Could navigation tools help users search and explore videos? In: Proceedings of the 2022 Conference on Human Information...

  37. [45]

    In: Proc

    Piao, G.: Recommending knowledge concepts on mooc platforms with meta-path-based rep- resentation learning. In: Proc. of Int. Conf. on Educational Data Mining (2021)

  38. [46]

    In: Proceedings of the 2016 Conference on User Modeling Adaptation and Personalization

    Piao, G., Breslin, J.G.: Analyzing aggregated semantics-enabled user modeling on google+ and twitter for personalized link recommendations. In: Proceedings of the 2016 Conference on User Modeling Adaptation and Personalization. UMAP ’16 (2016)

  39. [47]

    In: Proceedings of the 2016 Conference on User Modeling Adaptation and Personalization

    Piao, G., Breslin, J.G.: Analyzing mooc entries of professionals on linkedin for user modeling and personalized mooc recommendations. In: Proceedings of the 2016 Conference on User Modeling Adaptation and Personalization. UMAP ’16 (2016)

  40. [48]

    In: Proc

    Piccinno, F., Ferragina, P.: From tagme to wat: A new entity annotator. In: Proc. of the First Int. Workshop on Entity Recognition & Disambiguation. ERD ’14 (2014)

  41. [49]

    In: Proceedings of the First International Workshop on Entity Recognition & Disambiguation

    Piccinno, F., Ferragina, P.: From tagme to wat: a new entity annotator. In: Proceedings of the First International Workshop on Entity Recognition & Disambiguation. p. 55–62. ERD ’14, Association for Computing Machinery (2014). https://doi.org/10.1145/2633211.2634350, https://d...

  42. [50]

    In: Advances in Neural Information Processing Systems (2015) 34 Bulathwela et al

    Piech, C., Bassen, J., Huang, J., Ganguli, S., Sahami, M., Guibas, L.J., Sohl-Dickstein, J.: Deep knowledge tracing. In: Advances in Neural Information Processing Systems (2015) 34 Bulathwela et al

  43. [51]

    In: International Conference on Web Engineering

    Piro, L., Bianchi, T., Alessandrelli, L., Chizzola, A., Casiraghi, D., Sancassani, S., Gatti, N.: Mylearningtalk: An llm-based intelligent tutoring system. In: International Conference on Web Engineering. pp. 428–431. Springer (2024)

  44. [53]

    Knowledge-Based Systems188 (2020)

    Ponza, M., Ferragina, P., Chakrabarti, S.: On computing entity relatedness in wikipedia, with applications. Knowledge-Based Systems188 (2020)

  45. [54]

    Proceedings of the AAAI Conference on Artificial Intelligence 38(21), 23128–23136 (2024)

    Qiu, Y., Djemili, K., Elezi, D., Shalman Srazali, A., P ´erez-Ortiz, M., Yilmaz, E., Shawe- Taylor, J., Bulathwela, S.: A toolbox for modelling engagement with educational videos. Proceedings of the AAAI Conference on Artificial Intelligence 38(21), 23128–23136 (2024). https:/...

  46. [55]

    Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J., et al.: Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res. 21(140), 1–67 (2020)

  47. [56]

    Rajpurkar, P., Zhang, J., Lopyrev, K., Liang, P.: Squad: 100,000+ questions for machine comprehension of text (2016), https://arxiv.org/abs/1606.05250

  48. [57]

    In: Proceedings of the 4th Conference on Conversational User Interfaces

    Reicherts, L., Park, G.W., Rogers, Y.: Extending chatbots to probe users: Enhancing complex decision-making through probing conversations. In: Proceedings of the 4th Conference on Conversational User Interfaces. CUI ’22, Association for Computing Machinery, New York, NY, USA (...

  49. [58]

    Social Media Tools and Platforms in Learning Environments pp

    Sampson, D.G., Zervas, P., Kalamatianos, A.: Ask-lost 2.0: A web-based tool for social tagging digital educational resources in learning environments. Social Media Tools and Platforms in Learning Environments pp. 387–398 (2011)

  50. [59]

    Advances in Neural Information Processing Systems36 (2024)

    Schick, T., Dwivedi-Yu, J., Dess`ı, R., Raileanu, R., Lomeli, M., Hambro, E., Zettlemoyer, L., Cancedda, N., Scialom, T.: Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems36 (2024)

  51. [60]

    arXiv preprint arXiv:2405.16533 (2024)

    Shi, Z., Gao, S., Chen, X., Feng, Y., Yan, L., Shi, H., Yin, D., Chen, Z., Verberne, S., Ren, Z.: Chain of tools: Large language model is an automatic multi-tool learner. arXiv preprint arXiv:2405.16533 (2024)

  52. [61]

    arXiv preprint arXiv:2403.03031 (2024)

    Shi, Z., Gao, S., Chen, X., Feng, Y., Yan, L., Shi, H., Yin, D., Ren, P., Verberne, S., Ren, Z.: Learning to use tools via cooperative and interactive agents. arXiv preprint arXiv:2403.03031 (2024)

  53. [62]

    In: LAK21: 11th International Learning Analytics and Knowledge Conference

    Shin, D., Shim, Y., Yu, H., Lee, S., Kim, B., Choi, Y.: Saint+: Integrating temporal features for ednet correctness prediction. In: LAK21: 11th International Learning Analytics and Knowledge Conference. pp. 490–496 (2021)

  54. [63]

    Transactions of the Association for Computational Linguistics 12, 1063–1086 (09 2024)

    Sourati, Z., Ilievski, F., Sommerauer, P., Jiang, Y.: ARN: Analogical Reasoning on Narratives. Transactions of the Association for Computational Linguistics 12, 1063–1086 (09 2024)

  55. [64]

    In: International Conference on Artificial Intelligence in Education

    Stamper, J., Xiao, R., Hou, X.: Enhancing llm-based feedback: Insights from intelligent tutoring systems and the learning sciences. In: International Conference on Artificial Intelligence in Education. pp. 32–43. Springer (2024)

  56. [65]

    Syed, R., Collins-Thompson, K.: Optimizing search results for human learning goals. Inf. Retr. J. 20(5), 506–523 (2017)

  57. [66]

    Frontiers in Psychology15, 1474892 (2024)

    Th¨ us, D., Malone, S., Br¨ unken, R.: Exploring generative ai in higher education: a rag system to enhance student engagement with scientific literature. Frontiers in Psychology15, 1474892 (2024)

  58. [67]

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

  59. [68]

    International Educational Data Mining Society (2021)

    Williamson, K., Kizilcec, R.F.: Effects of algorithmic transparency in bayesian knowledge tracing on trust and perceived accuracy. International Educational Data Mining Society (2021)

  60. [69]

    International Educational Data Mining Society (2016)

    Xiong, X., Zhao, S., Van Inwegen, E.G., Beck, J.E.: Going deeper with deep knowledge tracing. International Educational Data Mining Society (2016)

  61. [70]

    In: Lane, H.C., Yacef, K., Mostow, J., Pavlik, P

    Yudelson, M.V., Koedinger, K.R., Gordon, G.J.: Individualized bayesian knowledge tracing models. In: Lane, H.C., Yacef, K., Mostow, J., Pavlik, P. (eds.) Proc. of Artificial Intelligence in Education (2013) TrueReason: Integrating Reasoning with Foundational Models 35

  62. [71]

    Zarrinkalam, F., Faralli, S., Piao, G., Bagheri, E.: Extracting, mining and predicting users’ interests from social media (2020)

  63. [72]

    Zhang, R., Guo, J., Chen, L., Fan, Y., Cheng, X.: A review on question generation from natural language text. Trans. on Information Systems40(1), 1–43 (2021)

  64. [73]

    Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., Artzi, Y.: Bertscore: Evaluating text gener- ation with bert (2020), https://arxiv.org/abs/1904.09675

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.