Paper deep dive
TeachingCoach: A Fine-Tuned Scaffolding Chatbot for Instructional Guidance to Instructors
Isabel Molnar, Peiyu Li, Si Chen, Sugana Chawla, James Lang, Ronald Metoyer, Ting Hua, Nitesh V. Chawla
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 3/22/2026, 6:00:50 AM
Summary
TeachingCoach is a pedagogically grounded, fine-tuned chatbot designed to provide instructional guidance to higher education instructors. It utilizes a data-centric pipeline that extracts pedagogical rules from educational resources and employs synthetic dialogue generation via GPT-4o to fine-tune a LLaMA-2-13B-Chat model. The system supports instructors through structured, multi-turn conversations focused on problem identification, diagnosis, and strategy development, demonstrating superior performance in clarity and reflection compared to general-purpose LLM baselines.
Entities (5)
Relation Signals (3)
GPT-4o → generatesdatafor → TeachingCoach
confidence 100% · GPT-4o produces multi-turn conversations... The dataset serves as high-quality input for fine-tuning the model.
TeachingCoach → isbasedon → LLaMA-2-13B-Chat
confidence 100% · We fine-tune a LLaMA-2-13B-Chat model... deployed as the TeachingCoach chatbot.
TeachingCoach → supports → Higher Education Instructors
confidence 95% · designed to support instructor professional development through real-time, conversational guidance.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Higher education instructors often lack timely and pedagogically grounded support, as scalable instructional guidance remains limited and existing tools rely on generic chatbot advice or non-scalable teaching center human-human consultations. We present TeachingCoach, a pedagogically grounded chatbot designed to support instructor professional development through real-time, conversational guidance. TeachingCoach is built on a data-centric pipeline that extracts pedagogical rules from educational resources and uses synthetic dialogue generation to fine-tune a specialized language model that guides instructors through problem identification, diagnosis, and strategy development. Expert evaluations show TeachingCoach produces clearer, more reflective, and more responsive guidance than a GPT-4o mini baseline, while a user study with higher education instructors highlights trade-offs between conversational depth and interaction efficiency. Together, these results demonstrate that pedagogically grounded, synthetic data driven chatbots can improve instructional support and offer a scalable design approach for future instructional chatbot systems.
Tags
Links
- Source: https://arxiv.org/abs/2603.18189v1
- Canonical: https://arxiv.org/abs/2603.18189v1
Trouble viewing inline? Open PDF directly →
Full Text
26,957 characters extracted from source content.
Expand or collapse full text
TeachingCoach: A Fine-Tuned Scaffolding Chatbot for Instructional Guidance to Instructors Isabel Molnar ∗ imolnar@nd.edu University of Notre Dame Notre Dame, Indiana, USA Peiyu Li ∗ pli9@nd.edu University of Notre Dame Notre Dame, Indiana, USA Si Chen schen34@nd.edu University of Notre Dame Notre Dame, Indiana, USA Sugana Chawla schawla@nd.edu University of Notre Dame Notre Dame, Indiana, USA James Lang jlang4@nd.edu University of Notre Dame Notre Dame, Indiana, USA Ronald Metoyer rmetoyer@nd.edu University of Notre Dame Notre Dame, Indiana, USA Ting Hua thua@nd.edu University of Notre Dame Notre Dame, Indiana, USA Nitesh V. Chawla nchawla@nd.edu University of Notre Dame Notre Dame, Indiana, USA Abstract Higher education instructors often lack timely and pedagogically grounded support, as scalable instructional guidance remains lim- ited and existing tools rely on generic chatbot advice or non-scalable teaching center human-human consultations. We present Teach- ingCoach, a pedagogically grounded chatbot designed to support instructor professional development through real-time, conversa- tional guidance. TeachingCoach is built on a data-centric pipeline that extracts pedagogical rules from educational resources and uses synthetic dialogue generation to fine-tune a specialized language model that guides instructors through problem identification, diag- nosis, and strategy development. Expert evaluations show Teach- ingCoach produces clearer, more reflective, and more responsive guidance than a GPT-4o mini baseline, while a user study with higher education instructors highlights trade-offs between conver- sational depth and interaction efficiency. Together, these results demonstrate that pedagogically grounded, synthetic data driven chatbots can improve instructional support and offer a scalable design approach for future instructional chatbot systems. CCS Concepts • Human-centered computing→Human computer interac- tion (HCI);• Applied computing→Computer-assisted in- struction. ∗ Both authors contributed equally to this work. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. Learning@Scale ’26, Seoul, South Korea © 2026 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 978-x-x-x-x/Y/M https://doi.org/10.1145/n.n Keywords Instructional Support, Chatbot, Conversational AI, Dialogue Sys- tems ACM Reference Format: Isabel Molnar, Peiyu Li, Si Chen, Sugana Chawla, James Lang, Ronald Metoyer, Ting Hua, and Nitesh V. Chawla. 2026. TeachingCoach: A Fine- Tuned Scaffolding Chatbot for Instructional Guidance to Instructors. In Proceedings of ACM Learning at Scale conference (Learning@Scale ’26). ACM, New York, NY, USA, 5 pages. https://doi.org/10.1145/n.n 1 Introduction Effective teaching requires not only subject-matter expertise but also skillful use of instructional strategies. Although educational research provides abundant rules and principles, instructors often struggle to apply them systematically in classroom contexts. Uni- versities operate teaching and learning centers that offer workshops and consultations, but these resources face key limitations: (1) sup- port may not be available at the moment of need, (2) feedback is often generic rather than tailored to an instructor’s background, and (3) some educators hesitate to seek direct help for fear of ap- pearing unskilled. As a result, many lack scalable and accessible support for their professional growth. Prior research on intelligent tutoring systems [5,8] shows that dialogue can scaffold reasoning and reflection, approaching the effectiveness of human tutors. More recently, educational chatbots and LLM-based frameworks [4,11,16,21] have been developed to model or support students, but surveys note that these systems are often weakly grounded in pedagogy and rarely address instructor professional development [3, 13]. Recent work has focused on supporting instructors through AI- enabled analytics, assessment scalability, and workload reduction. For example, studies examined the use of AI-generated practice questions to scale established learning effects [19], conversational interfaces that help instructors interpret learning analytics [22], and LLM-based assignment report summaries that support teacher arXiv:2603.18189v1 [cs.AI] 18 Mar 2026 Learning@Scale ’26, June 30 – July 3, 2026, Seoul, South KoreaMolnar*, Li* et al. insight [10]. However, little work directly targets instructional sup- port grounded in pedagogical practices for higher education in- structors; instead, such support often relies on teaching center consultants, which is difficult to scale and frequently unavailable at smaller or less-resourced institutions, highly US-centric, where instructional support may be limited or entirely absent. 1 Meanwhile, general-purpose LLMs like ChatGPT have become widely accessible, but their responses are typically generic and rarely apply evidence-based principles in a pedagogically scaffolded way. To bridge this gap, we introduce TeachingCoach, a chatbot system that simulates the role of a teaching expert and delivers conversational guidance on instructional practice. At its core is a pipeline that grounds the system in evidence-based pedagogy while enabling scalable data generation and model training. Figure 1 illustrates this process, which we detail in the following sections. 2 Related Works Large language models (LLMs) have been widely used as data gen- erators to synthesize labeled examples and structured interactions for model training, motivated by the cost, scarcity, and privacy lim- itations of human-collected data, as well as its susceptibility to bias and annotation noise [6]. LLM-based data generation has been suc- cessfully applied across domains such as code generation [12] and instruction following [20], and has played a central role in training mainstream models including Alpaca [17]. TeachingCoach builds on this line of work by using LLMs to synthesize multi-turn instruc- tional dialogues and training on curated synthetic data filtered by experts. Beyond their role as data generators, recent work has also ex- plored the use of LLMs as interactive tutors and instructional agents for learning support. Recent work has examined LLMs for learning and tutoring, high- lighting both their promise and limitations. Evaluations such as EducationQ show that teaching effectiveness does not scale lin- early with model size or general reasoning ability [16]. Systems such as PACE and Agent4Edu incorporate pedagogical strategies through Socratic questioning and learner modeling to enable per- sonalized support [4,11]. However, these systems primarily focus on student-centered tutoring. In contrast, recent human-centered design research explores AI-based pedagogical agents that directly support instructors by fostering trust, social transparency, and flexible engagement, particularly to encourage adoption among AI-conservative educators [1]. Related work has also examined classroom uses of generative AI with an emphasis on AI literacy. Studies with middle school English Language Arts teachers show how scaffolded AI use can support students’ understanding of AI concepts and ethical use [14], but position instructors mainly as facilitators rather than recipients of instructional support. Similarly, recent approaches that train LLM-based tutors using student models and pedagogical rubrics prioritize student outcomes and offer limited support for instructors’ reflective practices or instructional decision-making [15]. 1 https://podnetwork.org/content/uploads/Wright_PNN_NoCTLs_Jan2019_ update2pdf.pdf 3 Pipeline of TeachingCoach 3.1 Rules Extraction We start by building a pedagogical knowledge base from authorita- tive education resources [2,9], where experts extract 36 core instruc- tional rules and encode them as structured system prompts. Each rule functions as a guiding pedagogical principle in dialogue gener- ation, ensuring that conversations remain anchored in evidence- based practices rather than drifting toward generic or unsupported advice. 3.2 Data Collection To support realistic and diverse training data, we use GPT-4o [7] to generate teacher profiles that specify years of experience and teaching subject, as well as teaching challenges such as managing classroom attention. Given the extracted rules, teacher profiles, and challenges, GPT-4o produces multi-turn conversations between a teacher and a simulated expert. These dialogues capture how instructional principles may be applied in practice. To guarantee quality, human experts then review the conversations, removing those that are inconsistent, repetitive, or pedagogically unsound. The dataset serves as high-quality input for fine-tuning the model. After filtering, the dataset contains 406,183 training, 4,156 valida- tion, and 4,143 test examples. Specifically, each multi-turn conversation is generated by start- ing from a single validated teaching dilemma that implicitly violates one of the 36 practices. This scenario is paired with a synthetic instructor profile (e.g., course type, class size, teaching experience, and personality cues) to ground the interaction in a realistic class- room context. The conversation then unfolds as a dialogue between the instructor and an expert teaching consultant over 20–30 turns, progressing through stages of clarifying the teaching challenge, exploring possible instructional strategies, planning concrete next steps, and reflecting on how those strategies might work in prac- tice. Instructor turns are conditioned on the scenario and profile to surface realistic concerns and follow-up questions, while expert responses are constrained to be concise, supportive, and actionable, mirroring authentic teaching-center consultations. 3.3 Pedagogically-Grounded Model Training We fine-tune a LLaMA-2-13B-Chat model [18] with full parameter update on curated multi-turn dialogues. Each dialogue consists of a system message (teaching rules for the current instructional phase), a user message (profile and challenge), and an assistant message that first predicts a step descriptor and then generates the response. Steps follow the Eberly Center strategies 2 –Step 1: Identify the Problem, Step 2: Explore Reasons, and Step 3: Develop Strategies. Here is an example: System: [rules keyed by phase] User: [profile + challenge] Assistant: <step>Step 1</step> [response] ... User: [follow-up] Assistant: <step>Step 2</step> [response] 2 https://w.cmu.edu/teaching/solveproblem/index.html. Work-in-ProgressLearning@Scale ’26, June 30 – July 3, 2026, Seoul, South Korea Figure 1: Pipeline of TeachingCoach. Teaching guidelines are extracted from education resources, while GPT-4o generates teacher profiles and teaching challenges. These inputs prompt GPT-4o to produce synthetic multi-turn conversations, which, after expert filtering, are used to fine-tune a LLaMA model deployed as the TeachingCoach chatbot. Let퐻 (푖) 푡 denote the full conversation history up to assistant turn 푡in dialogue푖(including all prior system/user/assistant turns). The assistant is trained to (1) identify the current instructional step 휎 (푖) 푡 from퐻 (푖) 푡 , and then (2) generate the response푦 (푖) 푡 conditioned on both퐻 (푖) 푡 and휎 (푖) 푡 . We optimize the next-token objective over the entire assistant turn (step tokens + response tokens): L(휃)=− ∑︁ 푖,푡 log푃 휃 [휎 (푖) 푡 ,푦 (푖) 푡 ] | 퐻 (푖) 푡 , At inference, the model first generates<step>Step 푚</step> by leveraging퐻 푡 (the dialogue so far), thereby judging which in- structional stage applies; it then produces a response guided by the corresponding step-specific rules provided in the system prompt. For the end-user interface, the step tag can be logged for analytics and personalization while optionally being hidden from the visible transcript. Because the step label itself is generated by the model, Teaching- Coach aligns dialogue progression with a structured pedagogical scaffold, producing conversations that naturally advance from prob- lem identification to diagnosis to strategy development. We chose LLaMA-13B as the backbone model because it provides sufficient capability to model structured, multi-turn instructional dialogue while remaining feasible to fine-tune on academic com- pute resources. Its open-weight nature enables full supervision over both instructional step predictions and response generation, which is essential for implementing pedagogically grounded training ob- jectives. We compare against GPT-4o as a baseline because it represents a strong, widely deployed general-purpose LLM. Evaluating GPT-4o in a zero-shot setting highlights the impact of explicit pedagogical supervision and step-aware training in TeachingCoach, indepen- dent of model scale or proprietary data. 4 Evaluation We asked teaching experts to evaluate 200 conversations generated by TeachingCoach and a GPT-4o baseline. Each dialogue turn was rated on four dimensions: E1 clarity of responses, E2 respectful tone, E3 encouragement of reflection and reasoning, and E4 acknowledg- ment of user input. Ratings used a 3-point scale (1 = poor, 3 = strong). Table 1 summarizes the results, showing that TeachingCoach con- sistently outperformed GPT-4o. E1 E2 E3 E4 GPT-4o baseline2.011.991.681.85 TeachingCoach 2.62 2.55 2.59 2.56 Table 1: Average expert ratings over 200 conversations (3- point scale). Higher scores are marked in bold. Figure 2: The onboarding asks users to specific their experi- ence, current courses, and AI attitudes Figure 3: The chatbot interface allows users to interact with the TeachingCoach model, demonstrating multi-turn instruc- tional support. Learning@Scale ’26, June 30 – July 3, 2026, Seoul, South KoreaMolnar*, Li* et al. Figure 4: The dashboard provides a window for users to sched- ule consultations with live experts, view the resources they have collected, and manage their stored data 5 Demo System TeachingCoach provides three components: onboarding, the chat- bot interface, and the dashboard. Onboarding collects lightweight background information (e.g., experience, courses taught, key chal- lenges), to enable personalization with minimal user burden. The chatbot interface is the core feature, supporting multi-turn con- versations about teaching challenges. Dialogues follow a struc- tured flow—problem identification, cause exploration, and strategy development—enabling targeted, pedagogically grounded sugges- tions. As shown in Figure 3, an instructor might raise a concern such as student engagement, and the chatbot responds with strate- gies like low-stakes quizzes or technology-enabled check-ins. The dashboard acts as a hub for reflection and resource management, storing conversation summaries, generated suggestions, and user data. Together, these components make TeachingCoach accessible, personalized, and scalable for professional support. 6 Preliminary User Study and Results 6.1 Study Design We conducted a remote user study in August 2025 in which partici- pants interacted with two versions of a demo system: one powered by TeachingCoach and one by a GPT-4o-mini baseline. Participants explored the same teaching challenge with each system for up to 10 minutes per condition, with order counterbalanced. After each interaction, participants completed a brief survey and a 15-minute post-study interview. We recruited 41 higher education instructors through public postings across U.S. research universities, community colleges, and liberal arts colleges. The sample included 22 faculty members and 19 teaching assistants or instructional staff. Participants were informed that two versions of the system existed but were not told how the systems differed or which model powered each version. The session concluded with a semi-structured interview focused on perceived differences between systems, preferences, and perceived learning support. 6.2 Findings The results show a distinction between overall preference and per- ceived recall. While the baseline model was preferred overall (21 vs. 13), participants more often attributed learning to the fine-tuned model (18 vs. 13). Agreement between preference and learning attri- bution was low (39%), indicating that features driving preference do not always match perceived instructional value. Participants con- sistently distinguished the models on conversational engagement, efficiency of interaction, and breadth of suggestions. Conversational Engagement. The fine-tuned model was often described as engaging in dialogue resembling interaction with a human pedagogy expert. Participants noted its use of reflective questions and follow-up prompts that encouraged examination of teaching practices. This conversational depth supported learning through reflection, though some participants found it slowed the interaction when seeking quick answers. Efficiency of Interaction. Efficiency was a key strength of the baseline model. Participants valued its ability to generate responses quickly, enabling rapid access to ideas when seeking immediate guidance. In contrast, the fine-tuned model often required more conversational turns to reach concrete suggestions, which some perceived as inefficient, though others found the slower pace helpful for clarifying goals and context. Breadth of Suggestions. The baseline model was associated with greater breadth, offering wide-ranging strategies across contexts, which allowed users to scan for relevant ideas but several partici- pants felt overwhelmed. By contrast, the fine-tuned model offered a narrower, more focused set of suggestions centered on specific classroom practices, improving coherence while reducing exposure to diverse strategies. Overall the findings highlight a trade-off between depth and effi- ciency: the fine-tuned model supported learning through sustained engagement and reflection, while the baseline model prioritized speed and breadth of suggestions. 7 Discussion and Conclusion TeachingCoach demonstrates how LLMs can be trained to follow explicit instructional guidance by encoding instructional steps di- rectly into training dialogues. Rather than relying on large-scale human-collected teaching conversations, we use fully synthetic data generated from instructional guidelines, instructor profiles, and teaching challenges, enabling scalable training while avoiding privacy and ethical concerns. To understand how this step-aware instructional behavior is perceived in practice, we conducted a user study comparing Teach- ingCoach with a GPT-4o-mini baseline, which represents a widely used general-purpose conversational model optimized for efficiency and broad coverage. The study reveals a clear trade-off between interaction efficiency and instructional depth: while the baseline model was often valued for its speed and breadth of suggestions, TeachingCoach was more frequently associated with deeper instruc- tional support through reflective dialogue and structured reasoning. Rather than viewing this trade-off as a limitation, it motivates the development of hybrid instructional systems that balance breadth and depth based on user needs and context. One promising direc- tion is a multi-agent architecture, in which agents are specialized Work-in-ProgressLearning@Scale ’26, June 30 – July 3, 2026, Seoul, South Korea for complementary roles—for example, one prioritizing rapid idea generation and broad strategy exploration, and another providing structured instructional guidance and reflective questioning. Hybrid behavior can also be achieved within a single agent through adap- tive interaction modes, such as beginning with concise, breadth- oriented suggestions and transitioning into deeper instructional guidance when users seek reflection or clarification. In addition, explicit control options (e.g., quick suggestions versus guided re- flection) or progressive disclosure strategies can allow instructors to align system behavior with their immediate instructional goals. This study has several limitations. The sample comprised self- selected higher education instructors, which may limit generaliz- ability to K-12 settings, informal learning environments, or insti- tutions with different cultural and technological infrastructures. Additionally, the evaluation focused on instructors’ perceptions and did not directly assess student learning outcomes; future work should examine the effects of step-aware instructional dialogue on classroom practice and learning over time. Although our study focuses on higher education instructors, the data generation pipeline and training formulation are not limited to this context. By modifying instructor profiles, teaching challenges, and contextual constraints, the same approach can be applied to other instructor-facing settings, such as K–12 teaching, professional training, instructional coaching, mentoring, and tutoring support. More broadly, this work points toward a general approach for devel- oping instructor-facing instructional dialogue systems that support reflective teaching practice across domains, while preserving peda- gogical consistency, scalability, and privacy. Please find conversation examples and our codebase at https://osf. io/n2xyu/overview?view_only=e5b85d85b4a842dea902d9714f6faa67. References [1]Si Chen, Reid Metoyer, Khiem Le, Adam Acunin, Izzy Molnar, Alex Ambrose, James Lang, Nitesh Chawla, and Ronald Metoyer. 2025. Bridging the AI adoption gap: Designing an interactive pedagogical agent for higher education instructors. In International Conference on Artificial Intelligence in Education. Springer, 171– 178. [2]Flower Darby and James M Lang. 2019. Small teaching online: Applying learning science in online classes. John Wiley & Sons, San Fransisco, CA. [3]Tim Debets, Seyyed Kazem Banihashem, Desirée Joosten-Ten Brinke, Tanja EJ Vos, Gideon Maillette de Buy Wenniger, and Gino Camp. 2025. Chatbots in education: A systematic review of objectives, underlying technology and theory, evaluation criteria, and impacts. Computers & Education (2025), 105323. [4] Weibo Gao, Qi Liu, Linan Yue, Fangzhou Yao, Rui Lv, Zheng Zhang, Hao Wang, and Zhenya Huang. 2025. Agent4edu: Generating learner response data by generative agents for intelligent education systems. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 23923–23932. [5]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, 2 (2004), 180–192. [6]Or Honovich, Thomas Scialom, Omer Levy, and Timo Schick. 2023. Unnatural instructions: Tuning language models with (almost) no human labor. In Proceed- ings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 14409–14428. [7]Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al.2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276 (2024). [8]W Lewis Johnson and James C Lester. 2018. Pedagogical agents: back to the future. AI Magazine 39, 2 (2018), 33–44. [9] James M Lang. 2021. Small teaching: Everyday lessons from the science of learning. John Wiley & Sons, San Fransisco, CA. [10]Wen-Chiang Ivan Lim, Neil T Heffernan I, Ivan Eroshenko, Wai Khumwang, and Pei-Chen Chan. 2025. Leveraging LLMs for assignment report summaries to support teacher insights in intelligent tutoring systems. In Proceedings of the Twelfth ACM Conference on Learning@ Scale. 356–360. [11]Ben Liu, Jihai Zhang, Fangquan Lin, Xu Jia, and Min Peng. 2025. One size doesn’t fit all: A personalized conversational tutoring agent for mathematics instruction. In Companion Proceedings of the ACM on Web Conference 2025. 2401–2410. [12] Ziyang Luo, Can Xu, Pu Zhao, Qingfeng Sun, Xiubo Geng, Wenxiang Hu, Chongyang Tao, Jing Ma, Qingwei Lin, and Daxin Jiang. 2023. Wizardcoder: Empowering code large language models with evol-instruct. arXiv preprint arXiv:2306.08568 (2023). [13]Chinedu Wilfred Okonkwo and Abejide Ade-Ibijola. 2021. Chatbots applications in education: A systematic review. Computers and Education: Artificial Intelligence 2 (2021), 100033. [14]Daniel Ritchie, Tamara Tate, Yi Zhang, Kristi Werry, and Mark Warschauer. 2025. Supporting Middle School English Teachers’ AI Literacy Goals Through a Generative AI Tutor. In International Conference on Artificial Intelligence in Education. Springer, 283–290. [15]Alexander Scarlatos, Naiming Liu, Jaewook Lee, Richard Baraniuk, and Andrew Lan. 2025. Training LLM-Based Tutors to Improve Student Learning Outcomes in Dialogues. In International Conference on Artificial Intelligence in Education. Springer, 251–266. [16]Yao Shi, Rongkeng Liang, and Yong Xu. 2025. EducationQ: Evaluating LLMs’ Teaching Capabilities Through Multi-Agent Dialogue Framework. arXiv preprint arXiv:2504.14928 (2025). [17] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. 2023. Stanford alpaca: An instruction-following llama model. [18]Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yas- mine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhos- ale, et al.2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023). [19]Rachel Van Campenhout, Kevin S Autry, Michelle W Clark, and Benny G Johnson. 2025. Scaling the doer effect: A replication analysis using AI-generated questions. In Proceedings of the Twelfth ACM Conference on Learning@ Scale. 24–34. [20] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. Self-instruct: Aligning language mod- els with self-generated instructions. In Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers). 13484–13508. [21]Songlin Xu, Xinyu Zhang, and Lianhui Qin. 2024. Eduagent: Generative student agents in learning. arXiv preprint arXiv:2404.07963 (2024). [22]Chao Yang, Deliang Wang, and Gaowei Chen. 2025. Chat-LAD: Enhancing teacher understanding of learning analytics dashboard with AI-empowered explanations. In Proceedings of the Twelfth ACM Conference on Learning@ Scale. 197–201.