Pith. sign in

REVIEW 3 major objections 7 minor 46 references

Interactive Sketchpad: A Multimodal Tutoring System for Collaborative, Visual Problem-Solving

T0 review · 3 major / 7 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Interactive Sketchpad claims that a multimodal tutor which generates diagrams by executing code and lets students annotate them improves math comprehension, accuracy, and engagement over text-only chatbot tutoring.

desk verdict A sensible tutoring prototype whose headline learning claims are not backed by the reported evidence: seven self-reports are not accuracy data. read the letter →

arxiv 2503.16434 v2 pith:FM7HXUME submitted 2025-02-12 cs.HC cs.AI

Steven-Shine Chen , Jimin Lee , Paul Pu Liang This is my paper · ORCID
classification cs.HCcs.AI
keywords AIforeducationmultimodalinteractionvision-languagemodelsintelligenttutoringcodegenerationinteractivewhiteboardvisualreasoningmathematicslearning
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

The paper tries to establish that a tutoring system which generates supporting diagrams by writing and executing Python code, ties each diagram to a short textual hint, and lets students annotate the diagram on a whiteboard produces better math learning than a text-only chatbot. The system, Interactive Sketchpad, uses a pretrained multimodal model fine-tuned on human demonstrations of good hints, and it outperforms a non-interactive visual chain-of-thought baseline on IsoBench graph and function problems when allowed to use code execution. In a seven-participant user study, students rated the interface intuitive and the visuals accurate, and their comments tied the diagrams to conceptual understanding and the hint loop to a better learning experience. The practical stake is a reusable design pattern for educational AI: accurate diagrams computed by code, not hallucinated pixels, plus a collaborative annotation loop.

What carries the argument

The load-bearing mechanism is the code-generates-diagram loop. Instead of asking the model to draw pixels, the system prompts a fine-tuned GPT-4o to write a Python/matplotlib program; the program is executed by the model's code interpreter, and only the rendered image is shown to the student. This turns diagram accuracy into a programming problem, so coordinates, radii, and intersections are computed rather than guessed, and failed code can be retried until it runs. Around that loop sits the interactive whiteboard: the rendered diagram appears on a canvas, the student annotates it, and a screenshot is sent back to the model as multimodal input for the next hint.

What would settle it

Run a controlled learning study with a larger and more diverse sample: randomly assign students to Interactive Sketchpad or a text-only LLM tutor, score their solutions to comparable math problems before and after, and compare objective gain scores. The paper's claim fails if the multimodal condition does not produce a statistically meaningful accuracy gain over the text-only condition.

Watch

Extended reading notes

Core claim

The paper's central claim is that large multimodal models can be turned into effective visual tutors by adding an executable-code diagram step and an interactive whiteboard to the standard hint-giving loop. Concretely, Interactive Sketchpad first prompts a pretrained multimodal model to decide whether a visualization would help, then generates a matplotlib program that renders the diagram when run, then writes a hint that references the rendered image, and finally sends the image to a whiteboard where the student can annotate it and send the annotated screenshot back into the model. The paper reports that this system reaches 100% accuracy on IsoBench's max-flow tasks and outperforms Visual Sketchpad across graph and function families, and that users experienced improved comprehension, accuracy, and engagement. The improvements are framed as coming from code execution for exact calculations and diagrams plus Socratic step-by-step hints delivered in both text and vision.

Load-bearing premise

The learning claim rests on self-reported Likert ratings and qualitative comments from seven university students, with no before-and-after test, no objective measure of problem-solving accuracy, and no text-only control condition.

Editorial extensions

If this is right

  • On IsoBench tasks, the direct-solving version of Interactive Sketchpad beats Visual Sketchpad and plain GPT-4o on every task family, with the largest gains on max-flow (from 25.0 to 100.0) and isomorphism (from 50.8 to 75.0).
  • Because diagrams are produced by executing code, the system can compute coordinates, radii, and intersections exactly rather than trusting the model to draw from memory, which is what makes the visuals accurate enough to tutor with.
  • The interactive loop of hint, annotate, and send the screenshot back keeps the model from revealing the final answer, which participants said was better than ChatGPT-style direct answering.
  • The same four-step architecture of analyze, generate, hint, and annotate can be reused for any visually rich problem since the visualization path is Python code rather than task-specific image output.

Reading between the lines

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

  • A pre/post controlled trial with objective accuracy scoring is the natural next step; the current seven-user Likert evidence cannot separate the multimodal loop from novelty, the whiteboard hardware, or the hint-only policy.
  • The system's diagram failures, such as the hard-coded radius example, suggest a concrete extension: add a vision-based verifier that checks the rendered image against the problem facts and regenerates when misaligned.
  • The same code-generation-plus-annotation loop should transfer to physics and engineering tutoring, where force diagrams and trajectories play the role auxiliary lines play in geometry.
  • Student whiteboard annotations plus the model's diagram-generation traces could become training data for fine-tuning educational multimodal models to produce better hints.
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 / 7 minor

Summary. The paper presents Interactive Sketchpad, a tutoring system built on GPT-4o that generates code-drawn diagrams and textual hints for math problems, with an interactive whiteboard for annotation. The authors claim the system improves task comprehension, problem-solving accuracy, and engagement, based on a user study with seven university students and an IsoBench benchmark comparison. The technical architecture is described at system level, but the evaluation is limited to subjective self-reports and a non-equivalent baseline comparison.

Significance. If substantiated, a multimodal tutoring system that combines code-generated diagrams with textual hints and interactive whiteboard annotation would be a valuable contribution to HCI and AI-in-education, particularly because recent work warns that direct LLM answers can harm learning. The manuscript is timely and the prototype is clearly motivated. However, the evidence provided is far below what is needed to support the headline claims: the user study has no control comparison, no objective learning measure, no inferential statistics, and only seven participants; the only quantitative benchmark does not compare like with like. The code release and qualitative feedback are useful resources, but the central empirical claims are not established.

major comments (3)
  1. [§4.3.1, §4.3.4, Table 3] The abstract states that Interactive Sketchpad 'leads to improved task comprehension, problem-solving accuracy, and engagement levels,' but the only user-study evidence is self-reported Likert ratings from seven users (Table 3) and qualitative quotes (Table 2). No pre/post test, solution-accuracy measure, learning-gain score, control condition, or statistical test is reported, so the data do not support the claim of improved comprehension or accuracy; subjective 'effectiveness of hints' is not a measure of problem-solving accuracy.
  2. [§4.2, Table 1] The IsoBench comparison does not isolate the contribution of Interactive Sketchpad: the paper states that Visual Sketchpad's prompts constrain the model to avoid code execution, while Interactive Sketchpad allows it. The observed improvement could therefore be due solely to code execution rather than the interactive sketchpad/tutoring loop, making the comparison not apples-to-apples. Moreover, Table 1 measures direct problem solving by the system, not the effectiveness of the tutoring interaction that the user study is about.
  3. [§5, Table 4] The abstract also claims that 'accurate and robust diagrams' are generated, but Section 5 documents that visualization can fail and Table 4 shows a concrete incorrect diagram with hard-coded coordinates. At a minimum, the claim should be qualified, and the paper would need a systematic evaluation of diagram correctness (e.g., code-execution success rate or human judgment of diagram accuracy) to support it.
minor comments (7)
  1. [§4.3.1] The participant description is ungrammatical: 'university underclassmen undergraduate students participants' should be rephrased, and the enrollment and compensation details are incomplete.
  2. [§3 (Hint Generation)] The fine-tuning described in 'Hint Generation' gives no dataset size, training procedure, validation, or comparison to the non-fine-tuned model; this is important for reproducibility.
  3. [§4.3.4, Table 3] Report standard deviations or per-item distributions for Table 3; average scores alone cannot indicate agreement among participants.
  4. [§4.3.1] The user study appears to have been conducted without an approved IRB protocol, as the text states that approval was submitted only for 'further user studies'; this should be clarified or addressed with the venue.
  5. [§4.3.2] The citation [28] appears in the middle of the description of the user-study goal, but it would be clearer to place it with the specific claim about visualization benefits in mathematics learning and to separate it from the sentence about 'The primary goal...'.
  6. [§3, §4] The system is described as 'fine-tuned' in the abstract and introduction, but the experiments do not evaluate the fine-tuned model separately; please clarify what exactly was fine-tuned and whether the user study used the fine-tuned or prompted version.
  7. [Table 4] The text says the correct visualization is green and the incorrect one is red, but the table's images are not described in enough detail for readers to verify; please provide a text summary or accessible labels.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical, with no fitted parameters or derivations that reduce to their inputs.

full rationale

The paper does not present a derivation chain, fitted parameters, or mathematical predictions. Its central contribution is a tutoring system built on GPT-4o with code-executed visualizations and an interactive whiteboard, evaluated through an IsoBench performance comparison and a small user study. The IsoBench results (Table 1) are direct measured accuracies of the system on benchmark tasks; they are not predicted from fitted values. The user study reports qualitative feedback (Table 2) and average Likert ratings (Table 3), which support only subjective perceptions, not objective measures of task comprehension or problem-solving accuracy. This is an evidentiary limitation, not circularity: self-reported ratings are not equivalent by construction to the outcome claims in the abstract. The only self-citation is Ref. [15] (Liang et al.), used in the related-work discussion of multimodal learning; it is peripheral and not load-bearing for any result. The Limitations section's acknowledgment of occasionally incorrect visualizations and short-term metrics also does not reveal any circular step. Therefore the appropriate finding is no significant circularity.

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

The central claim depends on three unverified premises: that visual aids improve learning, that code-generated diagrams are accurate enough not to mislead, and that self-reported ratings measure actual learning gains.

assumptions (3)
  • domain assumption Visual aids improve mathematics learning.
    The paper motivates the system from cognitive science (refs [4], [22], [26], [39]) but does not test this premise; the user study assumes visual hints are beneficial.
  • domain assumption GPT-4o with Code Interpreter can generate accurate diagrams often enough to avoid confusing students.
    The system depends on code-generated diagrams being correct; Section 5 and Table 4 document failures where points are misaligned, and no error rate is reported.
  • ad hoc to paper Self-reported Likert ratings and qualitative comments measure task comprehension and problem-solving accuracy.
    Section 4.3.4 uses average user ratings as evidence for the abstract's claim of improved comprehension and accuracy, without direct pre/post measurement of task performance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Interactive Sketchpad: A Multimodal Tutoring System for Collaborative, Visual Problem-Solving." pith.science (2026). https://pith.science/paper/FM7HXUME

@misc{pith2026250316434,
  author       = {Pith},
  title        = {Pith review of: Interactive Sketchpad: A Multimodal Tutoring System for Collaborative, Visual Problem-Solving},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FM7HXUME}},
  note         = {Machine review of arXiv:2503.16434}
}
read the original abstract

Humans have long relied on visual aids like sketches and diagrams to support reasoning and problem-solving. Visual tools, like auxiliary lines in geometry or graphs in calculus, are essential for understanding complex ideas. However, many tutoring systems remain text-based, providing feedback only through natural language. Leveraging recent advances in Large Multimodal Models (LMMs), this paper introduces Interactive Sketchpad, a tutoring system that combines language-based explanations with interactive visualizations to enhance learning. Built on a pre-trained LMM, Interactive Sketchpad is fine-tuned to provide step-by-step guidance in both text and visuals, enabling natural multimodal interaction with the student. Accurate and robust diagrams are generated by incorporating code execution into the reasoning process. User studies conducted on math problems such as geometry, calculus, and trigonometry demonstrate that Interactive Sketchpad leads to improved task comprehension, problem-solving accuracy, and engagement levels, highlighting its potential for transforming educational technologies. All code is available at: https://stevenshinechen.github.io/interactivesketchpad/.

Figures

Figures reproduced from arXiv: 2503.16434 by the authors.

Figure 1
Figure 1. Interactive Sketchpad enhances GPT-4o’s ability to provide step-by-step, visual hints for problem-solving. Given a student query and problem statement, Interactive Sketchpad generates both textual hints and dynamic visual diagrams, allowing students to engage with the problem iteratively. Without Interactive Sketchpad, GPT-4o struggles to offer effective interactive guidance, frequently revealing the answer and not … view at source ↗
Figure 2
Figure 2. Overview of Interactive Sketchpad: Given a multimodal question, Interactive Sketchpad generates a program to create a visual aid, then uses the visual aid as part of a hint to help the user solve the problem. The visual aid is sent to the interactive whiteboard which the user can write and draw on before sending the annotated diagram back to receive feedback or further help. 1 INTRODUCTION Intelligent tutoring syste… view at source ↗
Figure 3
Figure 3. This flowchart demonstrates the interaction between user inputs, large multimodal models, and visualization [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Screenshot of the chatbot interface for Interactive Sketchpad. The user can view the generated visual hints and interact with Interac￾tive Sketchpad by typing messages and uploading images [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Screenshot of the interactive whiteboard [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: To check the value of an integral, Interac￾tive Sketchpad uses the trapezoidal rule to perform nu￾merical integration by writing a Python program using scipy.integrate [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: To solve for the side length of the base of a square [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 34 canonical work pages

  1. [1]

    Abraham Arcavi. 2003. The role of visual representations in the learning of mathematics. Educational studies in mathematics 52, 3 (2003), 215–241

    work page 2003

  2. [2]

    Rishabh Balse, Viraj Kumar, Prajish Prasad, and Jayakrishnan Madathil Warriem

    work page

  3. [3]

    Hamsa Bastani, Osbert Bastani, Alp Sungu, Haosen Ge, Ozge Kabakcı, and Rei Mariman. 2024. Generative ai can harm learning. A vailable at SSRN 4895486 (2024)

    work page 2024

  4. [4]

    Eliza Bobek and Barbara Tversky. 2016. Creating visual explanations improves learning. Cognitive research: principles and implications 1 (2016), 1–14

    work page 2016

  5. [5]

    Wenhu Chen, Xueguang Ma, Xinyi Wang, and William W. Cohen. 2023. Pro- gram of Thoughts Prompting: Disentangling Computation from Reasoning for Numerical Reasoning Tasks. Transactions on Machine Learning Research (2023). https://openreview.net/forum?id=YfZ4ZPt8zd

    work page 2023

  6. [6]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168 (2021)

    arXiv 2021

  7. [7]

    Deqing Fu, Ruohao Guo, Ghazal Khalighinejad, Ollie Liu, Bhuwan Dhingra, Dani Yogatama, Robin Jia, and Willie Neiswanger. 2024. IsoBench: Benchmarking Mul- timodal Foundation Models on Isomorphic Representations. In First Conference on Language Modeling. https://openreview.net/forum?id=KZd1EErRJ1

    work page 2024

  8. [8]

    Arthur C Graesser, Shulan Lu, George Tanner Jackson, Heather Hite Mitchell, Mathew Ventura, Andrew Olney, and Max M Louwerse. 2004. AutoTutor: A tutor with dialogue in natural language. Behavior Research Methods, Instruments, & Computers 36 (2004), 180–192

    work page 2004

Show all 46 references
  1. [9]

    Tanmay Gupta and Aniruddha Kembhavi. 2023. Visual programming: Com- positional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 14953–14962

    2023

  2. [10]

    Smith, and Ranjay Krishna

    Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A. Smith, and Ranjay Krishna. 2024. Visual Sketchpad: Sketching as a Visual Chain of Thought for Multimodal Language Models. In The Thirty-eighth Annual Conference on Neural Information Processin...

    2024

  3. [11]

    Majeed Kazemitabaar, Xinying Hou, Austin Henley, Barbara Jane Ericson, David Weintrop, and Tovi Grossman. 2023. How novices use LLM-based code generators to solve CS1 coding tasks in a self-paced learning environment. In Proceedings of the 23rd Koli Calling International Confe...

    2023

  4. [12]

    Sal Khan. 2023. Harnessing GPT-4 so that all students benefit. A nonprofit approach for equal access. Khan Academy Blog (2023)

    2023

  5. [13]

    Avila, Paul Lukowicz, Jochen Kuhn, Stefan Küchemann, and Jakob Karolus

    Lars Krupp, Steffen Steinert, Maximilian Kiefer-Emmanouilidis, Karina E. Avila, Paul Lukowicz, Jochen Kuhn, Stefan Küchemann, and Jakob Karolus. 2024. Chal- lenges and Opportunities of Moderating Usage of Large Language Models in Education. In Proceedings of the International ...

    2024

  6. [14]

    Hai Li, Chenglu Li, Wanli Xing, Sami Baral, and Neil Heffernan. 2024. Automated Feedback for Student Math Responses Based on Multi-Modality and Fine-Tuning. In Proceedings of the 14th Learning Analytics and Knowledge Conference . 763–770

    2024

  7. [15]

    Paul Pu Liang, Amir Zadeh, and Louis-Philippe Morency. 2024. Foundations & trends in multimodal machine learning: Principles, challenges, and open ques- tions. Comput. Surveys 56, 10 (2024), 1–42

    2024

  8. [16]

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2024. Let’s Verify Step by Step. In The Twelfth International Conference on Learning Representations. https://openreview.net/forum?...

    2024

  9. [17]

    Pan Lu, Ran Gong, Shibiao Jiang, Liang Qiu, Siyuan Huang, Xiaodan Liang, and Song-Chun Zhu. 2021. Inter-GPS: Interpretable Geometry Problem Solving with Formal Language and Symbolic Reasoning. In The 59th Annual Meeting of the Association for Computational Linguistics (ACL)

    2021

  10. [18]

    Liliana Mata, Gabriel Lazar, and Iuliana Lazar. 2016. Interactive Whiteboards for Teaching and Learning Science: Ascertaining Research. Online Submission 20, 2 (2016), 135–148

    2016

  11. [19]

    Allen Nie, Yash Chandak, Miroslav Suzara, Ali Malik, Juliette Woodrow, Matt Peng, Mehran Sahami, Emma Brunskill, and Chris Piech. 2024. The GPT Surprise: Offering Large Language Model Chat in a Massive Coding Class Reduced Engage- ment but Increased Adopters’ Exam Performances...

    2024

  12. [20]

    OpenAI. 2024. GPT-4o System Card. ArXiv abs/2410.21276 (2024). https: //api.semanticscholar.org/CorpusID:273662196

    2024 arXiv

  13. [21]

    Ethan Prihar, Morgan Lee, Mia Hopman, Adam Tauman Kalai, Sofia Vempala, Allison Wang, Gabriel Wickline, Aly Murray, and Neil Heffernan. 2023. Com- paring different approaches to generating mathematics explanations using large language models. In International Conference on Art...

    2023

  14. [22]

    Jamal Raiyn. 2016. The Role of Visual Learning in Improving Students’ High- Order Thinking Skills. Journal of Education and Practice 7, 24 (2016), 115–121

    2016

  15. [23]

    Daniel Philip Rose, Vaishnavi Himakunthala, Andy Ouyang, Ryan He, Alex Mei, Yujie Lu, Michael Saxon, Chinmay Sonar, Diba Mirza, and William Yang Wang

  16. [24]

    Bettina Rösken and Katrin Rolka. 2006. A picture is worth a 1000 words–the role of visualization in mathematics learning. In Proceedings 30th conference of the International Group for the Psychology of mathematics education , Vol. 4. Charles University Prague, 457–464

    2006

  17. [25]

    Johanna Schoenherr and Stanislaw Schukajlow. 2024. Characterizing exter- nal visualization in mathematics education research: a scoping review. ZDM– Mathematics Education 56, 1 (2024), 73–85

    2024

  18. [26]

    Ghulam Shabiralyani, Khuram Shahzad Hasan, Naqvi Hamad, and Nadeem Iqbal

  19. [27]

    Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. 2024. Visual CoT: Unleashing Chain-of-Thought Reasoning in Multi-Modal Language Models. CoRR abs/2403.16999 (2024). https: //doi.org/10.48550/arXiv.2403.16999

    2024 doi arXiv

  20. [28]

    Kyvete Shatri and Kastriot Buza. 2017. The Use of Visualization in Teaching and Learning Process for Developing Critical Thinking of Students. European Journal of Social Sciences Education and Research 9 (01 2017), 71. https://doi.org/10.26417/ ejser.v9i1.p71-74

    2017

  21. [29]

    Kumar Shridhar, Jakub Macina, Mennatallah El-Assady, Tanmay Sinha, Manu Ka- pur, and Mrinmaya Sachan. 2022. Automatic Generation of Socratic Subquestions for Teaching Math Word Problems. InProceedings of the 2022 Conference on Empir- ical Methods in Natural Language Processing...

    2022 doi

  22. [30]

    Dídac Surís, Sachit Menon, and Carl Vondrick. 2023. Vipergpt: Visual inference via python execution for reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 11888–11898

    2023

  23. [31]

    John Sweller. 1994. Cognitive load theory, learning difficulty, and instructional design. Learning and instruction 4, 4 (1994), 295–312

    1994

  24. [32]

    Kurt VanLehn. 2011. The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. Educational psychologist 46, 4 (2011), 197–221

    2011

  25. [33]

    Huanhuan Wang, Ahmed Tlili, Ronghuai Huang, Zhenyu Cai, Min Li, Zui Cheng, Dong Yang, Mengti Li, Xixian Zhu, and Cheng Fei. 2023. Examining the applica- tions of intelligent tutoring systems in real educational contexts: A systematic literature review from the social experimen...

    2023

  26. [34]

    Rose E Wang, Ana T Ribeiro, Carly D Robinson, Susanna Loeb, and Dora Demszky

  27. [35]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems 35 (2022), 24824–24837

    2022

  28. [36]

    Changrong Xiao, Wenxing Ma, Sean Xin Xu, Kunpeng Zhang, Yufang Wang, and Qi Fu. 2024. From Automation to Augmentation: Large Language Models Elevating Essay Scoring Landscape. arXiv preprint arXiv:2401.06431 (2024)

    2024 arXiv

  29. [37]

    Kevin P Yancey, Geoffrey Laflair, Anthony Verardi, and Jill Burstein. 2023. Rating short l2 essays on the cefr scale with gpt-4. InProceedings of the 18th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2023) . 576–584

    2023

  30. [38]

    arXiv preprint arXiv:2410.03017 (2024)

    Tutor CoPilot: A human-AI approach for scaling real-time expertise. arXiv preprint arXiv:2410.03017 (2024)

    2024 arXiv

  31. [39]

    Ying Zhang, Pengjin Wang, Wei Jia, Aijun Zhang, and Gaowei Chen. 2023. Dy- namic visualization by GeoGebra for mathematics learning: a meta-analysis of 20 years of research. Journal of Research on Technology in Education (2023), 1–22

    2023

  32. [40]

    point E

    Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. 2024. Trans- fusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039 (2024). Int...

    2024 arXiv

  33. [42]

    Jialu Zhang, José Pablo Cambronero, Sumit Gulwani, Vu Le, Ruzica Piskac, Gus- tavo Soares, and Gust Verbruggen. 2024. Pydex: Repairing bugs in introductory python assignments using llms. Proceedings of the ACM on Programming Lan- guages 8, OOPSLA1 (2024), 1100–1124

    2024

  34. [45]

    You should not give away the answer

    The generated actions can help the student with the user request # USER REQUEST #. You should not give away the answer

  35. [46]

    Equals ( LengthOf ( Line (Q , R ) ) , 6)

    You should be concise with your response Below are some examples of how to use the tools to solve the user requests . You can refer to them for help . You can also refer to the tool descriptions for more information . # EXAMPLE #: # USER REQUEST #: Given the geometry diagram <...

    2025

  36. [2015]

    Journal of education and practice 6, 19 (2015), 226–233

    Impact of visual aids in enhancing the learning process case research: District Dera Ghazi Khan. Journal of education and practice 6, 19 (2015), 226–233

    2015

  37. [2023]

    In Proceedings of the 16th Annual ACM India Compute Conference

    Evaluating the Quality of LLM-Generated Explanations for Logical Errors in CS1 Student Programs. In Proceedings of the 16th Annual ACM India Compute Conference. 49–54

  38. [2024]

    https://openreview.net/forum?id=01Yi8rzoNs

    Visual Chain of Thought: Bridging Logical Gaps with Multimodal Infillings. https://openreview.net/forum?id=01Yi8rzoNs

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.