pith. sign in

arxiv: 2606.11210 · v1 · pith:OEEKIMQZnew · submitted 2026-04-24 · 💻 cs.CL · cs.AI· cs.MM

T2MM: An LLM Supported Architecture For Inquiry-Based Modeling

John Kos , Rudra Singh , Ashok Goel This is my paper

Pith reviewed 2026-07-04 19:30 UTC · model glm-5.2

classification 💻 cs.CL cs.AIcs.MM
keywords LLM architectureinquiry-based learningmodel constructionaction sequence generationontology validationinteractive modelingscience educationVERA
0
0 comments X

The pith

Action sequences beat full code generation for LLM-driven model building

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper introduces Text to Multimodal Model (T2MM), an architecture that uses an LLM to generate structured action sequences (e.g., 'create node', 'edit relationship') rather than raw code, in order to help learners build and modify interactive ecological models in the VERA science learning platform. The core claim is that constraining the LLM's output to discrete actions grounded in the learner's current model state produces far more reliable model construction than asking the LLM to generate complete XML code, which is the dominant approach in prior work. T2MM achieves 100% compilation and 92.7% exact structural match against target models on a dataset of 975 procedurally generated learner requests, while baseline full-code-generation methods achieve 54.5% and 57.4% exact match. The paper argues that the action-sequence approach succeeds because it reduces the surface area for LLM hallucination, keeps the model editable between turns, and respects the ontology of the host system.

Core claim

The central finding is that an LLM asked to output a short list of validated actions on an existing model state is substantially more accurate than an LLM asked to output the entire model representation from scratch. On 975 test cases, T2MM compiled 100% of the time and produced structurally identical models 92.7% of the time, compared to 54.5% and 57.4% for zero-shot and few-shot full-code-generation baselines. The gap was most pronounced for model creation from an empty canvas, where the code-generation baselines failed almost entirely (0% and 8.3% success), while T2MM succeeded 79.4% of the time. The paper attributes this to the fact that action sequences are compact, can be checked and,,

What carries the argument

The mechanism is a four-stage pipeline: (1) the learner's current model is retrieved from the VERA database as XML and converted to a minimal JSON representation; (2) this JSON, the learner's natural language request, and a description of the VERA ontology are sent to the LLM (GPT-4.1-mini); (3) the LLM returns a list of structured actions such as 'add node', 'remove relationship', or 'update parameter', which are validated against the VERA ontology rules; (4) validated actions are mapped to VERA API calls that write changes to the model. The key contrast is with baselines that send the same prompt but ask the LLM to return complete XML, which introduces more tokens, more hallucination risk,

If this is right

  • Action-sequence architectures could generalize to other interactive learning environments where models must conform to a domain-specific ontology and remain editable across multiple learner turns.
  • The finding that full code generation fails on blank-canvas creation suggests that LLM-assisted tools relying on raw code output may be unsuitable for scaffolding the initial stages of open-ended modeling tasks.
  • The action-sequence approach could reduce cognitive load for learners by letting them describe desired changes in natural language while the system handles structural correctness, potentially freeing learners to focus on the scientific reasoning behind their models.
  • Ontology validation as a post-LLM checkpoint could become a standard design pattern for deploying LLMs in domains where outputs must satisfy formal constraints beyond what the LLM can guarantee on its own.

Load-bearing premise

The evaluation dataset is procedurally generated from five expert models using templated natural language phrases that closely mirror the system's action vocabulary (e.g., 'Add bear and owl, and make bear consume the owl'). This means the test requests are structurally simple and linguistically narrow. The authors acknowledge that real learner requests will likely be more ambiguous and varied, and the performance gap may narrow when the system faces genuine learner language.

What would settle it

Test T2MM against the baselines using natural language requests collected from actual learners using VERA, rather than procedurally generated templates. If the performance gap between T2MM and full-code-generation baselines narrows to statistical insignificance under real learner language, the central claim that action sequences are superior would be weakened.

Figures

Figures reproduced from arXiv: 2606.11210 by Ashok Goel, John Kos, Rudra Singh.

Figure 1
Figure 1. Figure 1: T2MM Architecture Diagram 2. T2MM: T2MM is further discussed in Section 3.2. In short, it requests the XML from the VERA database, converts it to JSON and sends that JSON, along with a prompt and the learner textual request to the LLM. T2MM then receives action sequences back from the LLM system which describe different actions that need to be taken on the current model to create the target model, e.g. Cre… view at source ↗
Figure 2
Figure 2. Figure 2: An VERA model representing the effects of hunting and logging on an ecosystem The front end presents the learner with a canvas with nodes representing biotics and abiotcs which they are able to drag and drop onto the canvas. Each biotic node has thirteen possible parameters representing descriptions such as: lifespan, carbon biomass, or number of offspring. Between any combination of biotic and abiotic nod… view at source ↗
Figure 3
Figure 3. Figure 3: Probability of Success by Length of target action sequence of success for subsequent attempts. The 0-Shot and N-Shot architectures were largely unsuccessful at creating a model when starting from an empty canvas, with the 0-Shot architecture failing in all cases to create a model. The N-Shot architecture was only slightly more successful successfully creating a model in 8.31% of cases. This gives some indi… view at source ↗
Figure 4
Figure 4. Figure 4: T2MM Success by action type of these systems for assistance in building a model from scratch, they would struggle to engage in inquiry as it would improperly represent their hypothesis, and would be unsuccessful in iteration. This also highlights the ability of T2MM in multi-step context aware model construction, which was a challenge outlined in the literature [7]. As for why our architecture is better ab… view at source ↗
read the original abstract

Model Construction is a foundational practice in science learning that relies on visualization and interactivity. Large Language Models, increasingly augmented with multimodal capabilities, have been integrated in education contexts to support learning. However, these tools lack visual interactivity that is required by some learning contexts. We introduce Text to Multimodal Model (T2MM), a robust, dynamic LLM supported architecture that assists in model construction within the open inquiry ecology-based modeling software Virtual Experimental Research Assistant (VERA). T2MM accounts for the current context of the learner's model and creates interactive models, rather than static images, enabling the model to remain responsive to manual adjustment. To measure technical feasibility, we evaluate T2MM through a custom procedurally generated dataset of natural language learner modeling requests and target models within the VERA system. T2MM outperforms a baseline model generation architecture implemented through LLM-supported full code generation, common in the literature, across all measured success metrics. Our contribution not only outlines LLM integration into a inquiry-based learning modeling tool, but also describes a possible architecture through which more interactive multimodal LLM tools can be created.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 7 minor

Summary. The paper introduces Text to Multimodal Model (T2MM), an LLM-supported architecture for interactive model construction within the VERA ecology modeling platform. Rather than having the LLM generate full XML representations of models, T2MM prompts the LLM to produce structured action sequences (e.g., create node, edit relationship) that are validated against the VERA ontology and applied via API calls to update the learner's model. The system is evaluated on a custom procedurally generated dataset of 975 natural language modeling requests paired with target models, against two baselines: 0-Shot and N-Shot full XML code generation. T2MM achieves 100% compilation and 92.7% exact structural match (graph edit distance = 0), compared to 54.5% and 57.4% for the baselines. The paper positions this as the first AI-in-Education architecture to use LLM-generated action sequences grounded in a live learner model state for interactive conceptual modeling.

Significance. The core architectural idea—using LLM-generated action sequences validated against an ontology rather than full code generation—is a reasonable and well-motivated design choice for interactive learning environments. The authors provide reproducible prompts and materials on OSF, which is commendable. The comparison against two code-generation baselines on four metrics provides useful evidence that the action-sequence approach yields higher structural fidelity. The system addresses a genuine gap: existing multimodal LLM architectures for education lack iterative, context-aware model construction. However, the significance of the evaluation is tempered by the procedurally generated nature of the dataset, which uses templated natural language closely tied to the action vocabulary, limiting the strength of claims about natural language understanding.

major comments (1)
  1. §3.3 (Dataset): The evaluation dataset is procedurally generated by decomposing expert models into intermediate states and pairing each transition with a templated natural language phrase tied to the action vocabulary (e.g., 'Add bear and owl, and make bear consume the owl'). The evaluation then measures whether T2MM can recover the target model from these phrases. This creates a risk of circularity: the NL input is generated from the same action vocabulary that T2MM expects as output, making the parsing task near-deterministic on simple cases. The paper acknowledges this limitation in §5 ('textual requests by real learners will present a unique challenge'), but the headline claims in the abstract and conclusion do not adequately contextualize that the 92.7% structural match is achieved on templated input that mirrors the system's action schema. The authors should either (a) reframe the
minor comments (7)
  1. §4.3, Table 3: The distribution of target action sequence lengths is heavily skewed: 600 of 975 cases (61.5%) have length 1, and some lengths (e.g., 4) have only 4 cases. Figure 3's success rates for lengths 4, 6, and 8 are based on very small samples. Consider adding sample sizes to Figure 3 or noting the small-n caveat in the caption.
  2. §4.2: The N-Shot baseline performs worse than 0-Shot (mean GED 3.572 vs. 1.638), which is counterintuitive. The paper attributes this to 'random parameter variation' but leaves it uninvestigated. While this does not affect T2MM's own results, it weakens the baseline comparison. A brief investigation or at least a more explicit acknowledgment that the N-Shot configuration may be suboptimal would strengthen the evaluation.
  3. §3.2: The JSON examples contain formatting artifacts (e.g., 'p r o p e r t i e s', 'so ur ce _i d', 'ta rg et _i d', 'r e m o v e _ r e l a t i o n s h i p'). These should be cleaned up for publication.
  4. §1: 'a inquiry-based learning modeling tool' should be 'an inquiry-based learning modeling tool'.
  5. §3.4: The graph edit distance metric is described qualitatively but not formally defined. A brief formal definition (or at least a clear statement of the cost model for node/edge/parameter edits) would improve reproducibility.
  6. §4.3: 'All architectures achieved full success for action sequence length of four, however there were only four cases in this category.' This sentence is somewhat misleading because it implies a strength, but n=4 provides no statistical power. Consider rephrasing to avoid overinterpretation.
  7. References: Reference [24] (Yu & Jiang, 2026) and [25] (Zheng et al., 2025) appear to be very recent or forthcoming; ensure citation details are complete and accurate.

Simulated Author's Rebuttal

1 responses · 0 unresolved

The referee acknowledges the core architectural contribution is well-motivated and the comparison against baselines is useful, but raises a major concern about the procedurally generated dataset: the templated NL input mirrors the system's action vocabulary, creating a risk of circularity that is not adequately reflected in the headline claims. The referee requests either reframing the claims or otherwise addressing this limitation more prominently.

read point-by-point responses

  1. Referee: §3.3 (Dataset): The evaluation dataset is procedurally generated by decomposing expert models into intermediate states and pairing each transition with a templated natural language phrase tied to the action vocabulary... This creates a risk of circularity: the NL input is generated from the same action vocabulary that T2MM expects as output, making the parsing task near-deterministic on simple cases. The paper acknowledges this limitation in §5 but the headline claims in the abstract and conclusion do not adequately contextualize that the 92.7% structural match is achieved on templated input that mirrors the system's action schema. The authors should either (a) reframe the claims or otherwise address this.

    Authors: We thank the referee for this thoughtful and accurate observation. The referee is correct that our dataset generation process creates a correspondence between the natural language phrasing and the action vocabulary, and that this limits the strength of claims we can make about natural language understanding. We agree that the abstract and conclusion do not sufficiently contextualize this limitation, and we will revise the manuscript accordingly. Specifically, we will: (1) Add an explicit caveat in the abstract noting that the evaluation uses procedurally generated, templated requests rather than natural learner language. (2) Revise the conclusion to state that the 92.7% structural match demonstrates technical feasibility of the action-sequence architecture on structured input, not robustness to the linguistic variability of real learner requests. (3) Expand the §5 limitations discussion to more directly acknowledge the circularity risk the referee identifies. We want to note two points of context, however. First, the dataset was designed intentionally to evaluate representational correctness and action sequencing — whether the architecture can reliably produce ontology-valid model edits — rather than NL understanding per se. The conclusion already states that 'this evaluation isolates representational correctness and action sequencing as prerequisites for learner-facing deployment, leaving linguistic variability and visual evaluation to future work,' but we agree this framing needs to be brought forward into the abstract. Second, the baselines (0-Shot and N-Shot) are evaluated on the identical dataset, so the relative comparison between action sequences and full code generation is not affected by the templating concern. The finding that action sequences yield higher ont revision: no

Circularity Check

1 steps flagged

Mild circularity in dataset construction: NL inputs are generated from the action vocabulary that T2MM is evaluated on recovering, but the paper is transparent about this and frames it as a feasibility test.

specific steps
  1. fitted input called prediction [Section 3.3 (Dataset), paragraph 2]
    "To generate the natural language request, we tied each of the actions described above to a natural language phase. For example, if the learner wanted to create a model consisting of a bear consuming an owl, the generated natural language request would be “Add bear and owl, and make bear consume the owl.” This gave us a sample of 975 model and natural language pairs."

    The evaluation input (NL request) is constructed by mapping each action to a fixed phrase. T2MM is then evaluated on whether it can recover the action sequence from the NL. Since the NL is generated FROM the action vocabulary that T2MM outputs, the task reduces to inverting a near-deterministic mapping. The 92.7% exact structural match is partly an artifact of this construction: the input encodes the output in templated form. However, the paper is transparent about this — it explicitly labels the dataset as 'procedurally generated,' acknowledges real learner requests 'will present a unique challenge,' and frames the evaluation as 'technical feasibility' rather than naturalistic performance. This is an evaluation validity concern more than a derivation that reduces to its inputs by formal定义

full rationale

The paper's central architectural contribution — generating structured action sequences rather than full XML code — has independent content and is not defined in terms of its evaluation output. The dataset construction does create a mild circularity: NL inputs are generated from the action vocabulary that T2MM is evaluated on recovering, making the inversion task near-deterministic on single-action cases (61.5% of the dataset). However, the paper is transparent about this construction, explicitly frames it as a feasibility test, and acknowledges the limitation with real learner language. Self-citations (refs [1], [20] involving Goel) are for background context, not load-bearing premises. No parameter is fitted to evaluation data and then 'predicted' back. The circularity is real but minor and honestly disclosed, warranting a score of 2.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 2 invented entities

The paper introduces no new mathematical constants or physical entities. The free parameters are prompt design choices hosted on OSF but not fully specified in the paper. The core axiom is that action sequences outperform full code generation for ontology-constrained tasks, which is supported by the evaluation but not independently established. The dataset generation method is ad hoc to this paper and its representativeness is the primary epistemic risk.

free parameters (2)
  • N (number of few-shot examples in N-Shot baseline) = Not explicitly stated; described as examples for node creation, edge creation, and parameter changes
    The number and content of few-shot examples for the N-Shot baseline are not fully specified in the paper, though prompts are hosted on OSF.
  • Prompt design for T2MM, 0-Shot, and N-Shot = Not specified in paper; hosted on OSF
    The system prompts, including the ontology description and few-shot examples, are design choices that affect performance but are not detailed in the paper itself.
axioms (4)
  • domain assumption LLM-generated action sequences are more reliable than LLM-generated full XML for ontology-constrained model construction.
    This is the core design assumption of T2MM, stated in Section 3.2 and discussed in Section 5. It is supported by the evaluation results but is not independently established outside this paper.
  • domain assumption Graph edit distance is an adequate proxy for model correctness in science learning contexts.
    Section 3.4 defines structural match as the evaluation metric. The paper acknowledges this only measures structural correctness, not visual or semantic appropriateness.
  • ad hoc to paper Procedurally generated templated requests are representative enough of learner behavior to demonstrate technical feasibility.
    The dataset generation process (Section 3.3) uses templated phrases tied to actions. The paper acknowledges this limitation in Section 5 but uses it as the basis for all claims.
  • domain assumption GPT 4.1-mini offers sufficient performance for this task at a suitable price point.
    Section 3 states this choice. Results may differ with other LLMs, but this is a reasonable engineering assumption.
invented entities (2)
  • T2MM Architecture independent evidence
    purpose: Converts natural language learner requests into validated action sequences for model construction in VERA
    The architecture is evaluated against baselines on a dataset of 975 examples. It makes falsifiable predictions about compilation and structural match rates.
  • Procedurally generated VERA dataset no independent evidence
    purpose: Evaluation dataset of 975 natural language request and target model pairs
    The dataset is custom-generated for this paper and is not independently validated against real learner behavior. Its representativeness is an open question acknowledged by the authors.

pith-pipeline@v1.1.0-glm · 13303 in / 2870 out tokens · 199612 ms · 2026-07-04T19:30:34.690337+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages

  1. [1]

    https://doi.org/10.48550/arXiv.2209.02576, http://arxiv.org/abs/2209.02576, arXiv:2209.02576 [cs]

    An, S., Bates, R., Rugaber, S., Hammock, J., Weigel, E., Goel, A.K.: Cognitive Assistance for Inquiry-Based Modeling (Sep 2022). https://doi.org/10.48550/arXiv.2209.02576, http://arxiv.org/abs/2209.02576, arXiv:2209.02576 [cs]

    work page doi:10.48550/arxiv.2209.02576 2022

  2. [2]

    Bauer, A., Flatten, J., Popovic, Z.: Analysis of Problem-Solving Behavior in Open- Ended Scientific-Discovery Game Challenges. Tech. rep., International Educational Data Mining Society (Jun 2017), https://eric.ed.gov/?id=ED596582, eRIC Num- ber: ED596582

    work page 2017

  3. [3]

    Learning and Individual Dif- ferences118, 102601 (Feb 2025)

    Bewersdorff, A., Hartmann, C., Hornberger, M., Seßler, K., Bannert, M., Kas- neci, E., Kasneci, G., Zhai, X., Nerdel, C.: Taking the Next Step with Generative Artificial Intelligence: The Transformative Role of Multimodal Large Language Models in Science Education. Learning and Individual Dif- ferences118, 102601 (Feb 2025). https://doi.org/10.1016/j.lind...

    work page doi:10.1016/j.lindif.2024.102601 2025

  4. [4]

    Applied Artificial Intelligence19(3- 4), 363–392 (Mar 2005)

    Biswas, G., Leelawong, K., Schwartz, D., Vye, N., The Teachable Agents Group At Vande: LEARNING BY TEACHING: A NEW AGENT PARADIGM FOR EDUCATIONAL SOFTWARE. Applied Artificial Intelligence19(3- 4), 363–392 (Mar 2005). https://doi.org/10.1080/08839510590910200, http://www.tandfonline.com/doi/abs/10.1080/08839510590910200

    work page doi:10.1080/08839510590910200 2005

  5. [5]

    Cognitive Research: Principles and Implications 1(1), 27 (Dec 2016)

    Bobek, E., Tversky, B.: Creating visual explanations im- proves learning. Cognitive Research: Principles and Implications 1(1), 27 (Dec 2016). https://doi.org/10.1186/s41235-016-0031-6, https://cognitiveresearchjournal.springeropen.com/articles/10.1186/s41235- 016-0031-6

    work page doi:10.1186/s41235-016-0031-6 2016

  6. [6]

    International Journal of Human-Computer Studies64(11), 1099–1114 (Nov 2006)

    Bridewell, W., Sánchez, J.N., Langley, P., Billman, D.: An interac- tive environment for the modeling and discovery of scientific knowl- edge. International Journal of Human-Computer Studies64(11), 1099–1114 (Nov 2006). https://doi.org/10.1016/j.ijhcs.2006.06.006, https://www.sciencedirect.com/science/article/pii/S1071581906001030

    work page doi:10.1016/j.ijhcs.2006.06.006 2006

  7. [7]

    https://doi.org/10.48550/arXiv.2407.17544, http://arxiv.org/abs/2407.17544, arXiv:2407.17544 [cs]

    Bulusu, A., Man, B., Jagmohan, A., Vempaty, A., Mari-Wyka, J., Akkil, D.: MathViz-E: A Case-study in Domain-Specialized Tool-Using Agents (Jul 2024). https://doi.org/10.48550/arXiv.2407.17544, http://arxiv.org/abs/2407.17544, arXiv:2407.17544 [cs]

    work page doi:10.48550/arxiv.2407.17544 2024

  8. [8]

    The Science Teacher62(7), 28 (Oct 1995),https://www.proquest.com/docview/214631067/abstract/63C7A5E3113D496BPQ/1, num Pages: 6

    Bybee, R.W.: Achieving scientific literacy. The Science Teacher62(7), 28 (Oct 1995),https://www.proquest.com/docview/214631067/abstract/63C7A5E3113D496BPQ/1, num Pages: 6

    work page arXiv 1995

  9. [9]

    Christie, S., Rafferty, A., Lee, Z., Cutler, E., Tian, Y., Almoubayyed, H.: An Agen- tic Framework for Real-Time Pedagogical Plot Generation. pp. 309–316 (Jul 2025). https://doi.org/10.1007/978-3-031-98462-4_39

    work page doi:10.1007/978-3-031-98462-4_39 2025

  10. [10]

    Journal of Science Teacher Education4(1), 1–8 (Mar 1993)

    Flick, L.B.: The meanings of hands-on science. Journal of Science Teacher Education4(1), 1–8 (Mar 1993). https://doi.org/10.1007/BF02628851, https://www.tandfonline.com/doi/full/10.1007/BF02628851

    work page doi:10.1007/bf02628851 1993

  11. [11]

    In: Cristea, A.I., Walker, E., Lu, Y., Santos, O.C., Isotani, S

    Gupta, A., Reddig, J., Calò, T., Weitekamp, D., MacLellan, C.: Beyond Final Answers: Evaluating Large Language Models for Math Tutoring. In: Cristea, A.I., Walker, E., Lu, Y., Santos, O.C., Isotani, S. (eds.) Artificial Intelli- gence in Education. pp. 323–337. Springer Nature Switzerland, Cham (2025). https://doi.org/10.1007/978-3-031-98414-3_23 14 John ...

    work page doi:10.1007/978-3-031-98414-3_23 2025

  12. [12]

    In: Cristea, A.I., Walker, E., Lu, Y., Santos, O.C., Isotani, S

    Hou, X., Forsyth, C., Andrews-Todd, J., Rice, J., Cai, Z., Jiang, Y., Zapata-Rivera, D., Graesser, A.: An LLM-Enhanced Multi-agent Architecture for Conversation- Based Assessment. In: Cristea, A.I., Walker, E., Lu, Y., Santos, O.C., Isotani, S. (eds.) Artificial Intelligence in Education. pp. 119–134. Springer Nature Switzer- land, Cham (2025). https://do...

    work page doi:10.1007/978-3-031-98417-4_9 2025

  13. [13]

    Educational Psycholo- gist41(2), 75–86 (Jun 2006)

    Kirschner, P.A., Sweller, J., Clark, R.E.: Why Minimal Guidance During Instruc- tion Does Not Work: An Analysis of the Failure of Constructivist, Discovery, Problem-Based, Experiential, and Inquiry-Based Teaching. Educational Psycholo- gist41(2), 75–86 (Jun 2006)

    work page 2006

  14. [14]

    Bibliovault OAI Repository, the University of Chicago Press87(Jan 2005)

    Kuhn, D.: Education for Thinking. Bibliovault OAI Repository, the University of Chicago Press87(Jan 2005)

    work page 2005

  15. [15]

    Computers in Human Behavior25(2), 284–289 (Mar 2009)

    Leutner, D., Leopold, C., Sumfleth, E.: Cognitive load and science text comprehen- sion: Effects of drawing and mentally imagining text content. Computers in Human Behavior25(2), 284–289 (Mar 2009). https://doi.org/10.1016/j.chb.2008.12.010, https://www.sciencedirect.com/science/article/pii/S0747563208002197

    work page doi:10.1016/j.chb.2008.12.010 2009

  16. [16]

    In: Cristea, A.I., Walker, E., Lu, Y., Santos, O.C., Isotani, S

    Li, W., Pea, R., Haber, N., Subramonyam, H.: CogGen: A Learner-Centered Gen- erative AI Architecture for Intelligent Tutoring with Programming Videos. In: Cristea, A.I., Walker, E., Lu, Y., Santos, O.C., Isotani, S. (eds.) Artificial In- telligence in Education. pp. 11–18. Springer Nature Switzerland, Cham (2025). https://doi.org/10.1007/978-3-031-98462-4_2

    work page doi:10.1007/978-3-031-98462-4_2 2025

  17. [17]

    Educational Psychologist38(1), 43–52 (Jan 2003)

    Mayer, R.E., Moreno, R.: Nine Ways to Reduce Cognitive Load in Multimedia Learning. Educational Psychologist38(1), 43–52 (Jan 2003)

    work page 2003

  18. [18]

    Chelsea Green Publishing (2008)

    Meadows, D.: Thinking in Systems: A Primer. Chelsea Green Publishing (2008)

    work page 2008

  19. [19]

    In: Cristea, A.I., Walker, E., Lu, Y., Santos, O.C., Isotani, S

    Mou,Y.,Fathi,F.,Thillen,B.,Decker,S.:WILLM:ASystemforAcademicWriting Improvement Based on Large Language Models. In: Cristea, A.I., Walker, E., Lu, Y., Santos, O.C., Isotani, S. (eds.) Artificial Intelligence in Education. pp. 36– 43.SpringerNatureSwitzerland,Cham(2025).https://doi.org/10.1007/978-3-031- 98462-4_5

    work page doi:10.1007/978-3-031- 2025

  20. [20]

    In: Cristea, A.I., Walker, E., Lu, Y., San- tos, O.C., Isotani, S

    Taneja, K., Singh, A., Goel, A.K.: Towards a Multimodal Document-Grounded Conversational AI System for Education. In: Cristea, A.I., Walker, E., Lu, Y., San- tos, O.C., Isotani, S. (eds.) Artificial Intelligence in Education. pp. 92–99. Springer Nature Switzerland, Cham (2025). https://doi.org/10.1007/978-3-031-98462-4_12

    work page doi:10.1007/978-3-031-98462-4_12 2025

  21. [21]

    Tisue, S., Wilensky, U.: NetLogo: A Simple Environment for Modeling Complexity

    work page

  22. [22]

    Interactive Learning Environments21(4), 371–413 (Aug 2013)

    VanLehn, K.: Model construction as a learning activity: a design space and review. Interactive Learning Environments21(4), 371–413 (Aug 2013). https://doi.org/10.1080/10494820.2013.803125

    work page doi:10.1080/10494820.2013.803125 2013

  23. [23]

    Cognition and Instruction24(2), 171–209 (Jun 2006)

    Wilensky, U., Reisman, K.: Thinking Like a Wolf, a Sheep, or a Firefly: Learning Biology Through Constructing and Testing Computational Theories—An Embod- ied Modeling Approach. Cognition and Instruction24(2), 171–209 (Jun 2006). https://doi.org/10.1207/s1532690xci2402_1

    work page doi:10.1207/s1532690xci2402_1 2006

  24. [24]

    https://doi.org/10.48550/arXiv.2601.05162, http://arxiv.org/abs/2601.05162, arXiv:2601.05162 [cs]

    Yu, J., Jiang, D.: GenAI-DrawIO-Creator: A Framework for Automated Diagram Generation (Jan 2026). https://doi.org/10.48550/arXiv.2601.05162, http://arxiv.org/abs/2601.05162, arXiv:2601.05162 [cs]

    work page doi:10.48550/arxiv.2601.05162 2026

  25. [25]

    https://doi.org/10.48550/arXiv.2510.20229, http://arxiv.org/abs/2510.20229, arXiv:2510.20229 [cs]

    Zheng, G., Qian, J., Tang, J., Yang, S.: Why LVLMs Are More Prone to Hallucinations in Longer Responses: The Role of Context (Oct 2025). https://doi.org/10.48550/arXiv.2510.20229, http://arxiv.org/abs/2510.20229, arXiv:2510.20229 [cs]

    work page doi:10.48550/arxiv.2510.20229 2025