Kenneth R. Koedinger

Jionghao Lin, Danielle R. Thomas, Feifei Han et al. · cmu

Research demonstrates learners engaging in the process of producing explanations to support their reasoning, can have a positive impact on learning. However, providing learners real-time explanatory feedback often presents challenges related to classification accuracy, particularly in domain-specific environments, containing situationally complex and nuanced responses. We present two approaches for supplying tutors real-time feedback within an online lesson on how to give students effective praise. This work-in-progress demonstrates considerable accuracy in binary classification for corrective feedback of effective, or effort-based (F1 score = 0.811), and ineffective, or outcome-based (F1 score = 0.350), praise responses. More notably, we introduce progress towards an enhanced approach of providing explanatory feedback using large language model-facilitated named entity recognition, which can provide tutors feedback, not only while engaging in lessons, but can potentially suggest real-time tutor moves. Future work involves leveraging large language models for data augmentation to improve accuracy, while also developing an explanatory feedback interface.

Dollaya Hirunyasiri, Danielle R. Thomas, Jionghao Lin et al. · cmu

Research suggests that providing specific and timely feedback to human tutors enhances their performance. However, it presents challenges due to the time-consuming nature of assessing tutor performance by human evaluators. Large language models, such as the AI-chatbot ChatGPT, hold potential for offering constructive feedback to tutors in practical settings. Nevertheless, the accuracy of AI-generated feedback remains uncertain, with scant research investigating the ability of models like ChatGPT to deliver effective feedback. In this work-in-progress, we evaluate 30 dialogues generated by GPT-4 in a tutor-student setting. We use two different prompting approaches, the zero-shot chain of thought and the few-shot chain of thought, to identify specific components of effective praise based on five criteria. These approaches are then compared to the results of human graders for accuracy. Our goal is to assess the extent to which GPT-4 can accurately identify each praise criterion. We found that both zero-shot and few-shot chain of thought approaches yield comparable results. GPT-4 performs moderately well in identifying instances when the tutor offers specific and immediate praise. However, GPT-4 underperforms in identifying the tutor's ability to deliver sincere praise, particularly in the zero-shot prompting scenario where examples of sincere tutor praise statements were not provided. Future work will focus on enhancing prompt engineering, developing a more general tutoring rubric, and evaluating our method using real-life tutoring dialogues.

Ashish Gurung, Jordan Gutterman, Danielle R. Thomas et al.

Measures of disengagement provide insights into unproductive use of learning opportunities. Although measures of active disengagement, such as gaming the system and mind-wandering, are well studied, loss of practice time due to outright task avoidance remains relatively understudied. The current study addresses this gap by extending existing within-task measures (idle time) with two new session-level measures (delayed start and early stop) to capture loss of practice time due to task avoidance. We characterize the combined lost time as coasted time and the associated behavior as coasting behavior. Using ASSISTments logs (N = 1,425), we find that students dedicate only 40% of available classwork time to math practice and coast through the remaining 60%. Of the coasted time, 36% resulted from delayed starts, 2% from mid-practice idling, and 62% from stopping early. Delayed start and early stop showed moderate temporal stability (G = 0.73 and 0.71, respectively), suggesting that coasting is a consistent behavioral pattern. Even after excluding early stops attributable to assignment completion (i.e., early stop = 0), coasted time remained substantial at 32%. While we observe significant differences in coasting by gender and IEP status, we do not observe them by other demographic factors or school locale. Critically, students who continued working beyond the first assignment completion ("extra effort") performed significantly better on standardized tests. For research, coasting offers a new lens on opportunity loss by combining session-level disengagement with within-task disengagement. For practitioners, our results highlight the need for platform affordances that support sustained engagement and more productive use of available practice time.

Corina Luca Focsan, Marie Cynthia Abijuru Kamikazi, Tamisha Thompson et al.

Accountable Talk theory has been widely adopted to analyze classroom discourse and is increasingly used to annotate tutoring interactions. In particular, the TalkMoves codebook, grounded in Accountable Talk theory, is commonly used to label tutoring data and train models of effective instructional support. However, Accountable Talk was originally developed to characterize collaborative, whole-classroom oral discourse, not to identify talk moves in one-on-one tutoring environments using multimodal data (e.g., video, audio, chat). As tutoring platforms expand in scale and modality, questions remain about whether Accountable Talk-based codebooks generalize reliably beyond their original classroom context and data representation. This study examines whether the human-developed TalkMoves codebook generalizes in reliability, utility, and interpretability when applied to one-on-one tutoring across audio, chat, and multimodal data. We compare TalkMoves with a hybrid AI-human developed codebook using a workflow established in prior research. Two expert annotators with over 20 years of teaching experience applied both codebooks to six tutoring sessions spanning three modalities: chat-based, audio-only, and multimodal interactions. Results show that while Talk-Moves achieved higher overall inter-rater reliability than the AI-human codebook (k = 0.74 vs. 0.64), the AI-human codebook demonstrated broader empirical coverage and higher perceived usability across modalities. Both codebooks undercaptured tutoring-relevant moves and introduced ambiguity when identifying actions expressed through nonverbal and multimodal artifacts. Together, these findings highlight the uneven generalizability of TalkMoves to tutoring contexts and motivate the development of modality-aware, tutoring-grounded codebooks.

Chloe Qianhui Zhao, Jie Cao, Jionghao Lin et al.

Providing timely, targeted, and multimodal feedback helps students quickly correct errors, build deep understanding and stay motivated, yet making it at scale remains a challenge. This study introduces a real-time AI-facilitated multimodal feedback system that integrates structured textual explanations with dynamic multimedia resources, including the retrieved most relevant slide page references and streaming AI audio narration. In an online crowdsourcing experiment, we compared this system against fixed business-as-usual feedback by educators across three dimensions: (1) learning effectiveness, (2) learner engagement, (3) perceived feedback quality and value. Results showed that AI multimodal feedback achieved learning gains equivalent to original educator feedback while significantly outperforming it on perceived clarity, specificity, conciseness, motivation, satisfaction, and reducing cognitive load, with comparable correctness, trust, and acceptance. Process logs revealed distinct engagement patterns: for multiple-choice questions, educator feedback encouraged more submissions; for open-ended questions, AI-facilitated targeted suggestions lowered revision barriers and promoted iterative improvement. These findings highlight the potential of AI multimodal feedback to provide scalable, real-time, and context-aware support that both reduces instructor workload and enhances student experience.

Danielle R. Thomas, Conrad Borchers, Kirk P. Vanacore et al.

Generative Artificial Intelligence (GenAI) is now widespread in education, yet the efficacy of GenAI systems remains constrained by the quality and interpretation of the labeled data used to train and evaluate them. Studies commonly report inter-rater reliability (IRR), often summarized by a single coefficient such as Cohen's kappa (k), as a gatekeeper to ``ground truth.'' We argue that many educational assessment and practice support settings include challenges, such as high-inference constructs, skewed label distributions, and temporally segmented multimodal data, which yield potential misapplication or misinterpretation of threshold-based heuristics for IRR. The growing use of large language models as annotators and judges introduces risks such as automation bias and circular validation. We propose four practical shifts for establishing ground truth: (1) treat IRR as a diagnostic signal to localize disagreement and refine constructs rather than a mechanical acceptance threshold (e.g., k > 0.8); (2) require transparent reporting of rater expertise, codebook development, reconciliation procedures, and segmentation rules; (3) mitigate risks in LLM annotation through bias audits and verification workflows; and (4) complement agreement statistics with validity and effectiveness evidence for the intended use, including uncertainty-aware labeling (e.g., assigning different labels to the same item to capture nuance), criterion-related checks (e.g., predictive tests to check if labels forecast the intended outcome), and close-the-loop evaluations of whether systems trained on these labels improve learning beyond a reasonable control. We illustrate these shifts through case studies of multimodal tutoring data and provide actionable recommendations toward strengthening the evidence base of labeled AIED datasets.

Canwen Wang, Jionghao Lin, Kenneth R. Koedinger

Knowledge Components (KCs) are foundational to adaptive learning systems, but their manual identification by domain experts is a significant bottleneck. While Large Language Models (LLMs) offer a promising avenue for automating this process, prior research has been limited to small datasets and has been shown to produce superfluous, redundant KC labels. This study addresses these limitations by first scaling a "simulated textbook" LLM prompting strategy (using GPT-4o-mini) to a larger dataset of 646 multiple-choice questions. We found that this initial automated approach performed significantly worse than an expert-designed KC model (RMSE 0.4285 vs. 0.4206) and generated an excessive number of KCs (569 vs. 101). To address the issue of redundancy, we proposed and evaluated a novel method for merging semantically similar KC labels based on their cosine similarity. This merging strategy significantly improved the model's performance; a model using a cosine similarity threshold of 0.8 achieved the best result, reducing the KC count to 428 and improving the RMSE to 0.4259. This demonstrates that while scaled LLM generation alone is insufficient, combining it with a semantic merging technique offers a viable path toward automating and refining KC identification.

Qianru Lyu, Conrad Borchers, Meng Xia et al.

Past research has defined a general process for the data-driven redesign of educational technologies and has shown that in carefully-selected instances, this process can help make systems more effective. In the current work, we test the generality of the approach by applying it to four units of a middle-school mathematics intelligent tutoring system that were selected not based on suitability for redesign, as in previous work, but on topic. We tested whether the redesigned system was more effective than the original in a classroom study with 123 students. Although the learning gains did not differ between the conditions, students who used the Redesigned Tutor had more productive time-on-task, a larger number of skills practiced, and greater total knowledge mastery. The findings highlight the promise of data-driven redesign even when applied to instructional units *not* selected as likely to yield improvement, as evidence of the generality and wide applicability of the method.

Zilyu Ji, Yuntian Shen, Jionghao Lin et al.

Protecting Personally Identifiable Information (PII), such as names, is a critical requirement in learning technologies to safeguard student and teacher privacy and maintain trust. Accurate PII detection is an essential step toward anonymizing sensitive information while preserving the utility of educational data. Motivated by recent advancements in artificial intelligence, our study investigates the GPT-4o-mini model as a cost-effective and efficient solution for PII detection tasks. We explore both prompting and fine-tuning approaches and compare GPT-4o-mini's performance against established frameworks, including Microsoft Presidio and Azure AI Language. Our evaluation on two public datasets, CRAPII and TSCC, demonstrates that the fine-tuned GPT-4o-mini model achieves superior performance, with a recall of 0.9589 on CRAPII. Additionally, fine-tuned GPT-4o-mini significantly improves precision scores (a threefold increase) while reducing computational costs to nearly one-tenth of those associated with Azure AI Language. Furthermore, our bias analysis reveals that the fine-tuned GPT-4o-mini model consistently delivers accurate results across diverse cultural backgrounds and genders. The generalizability analysis using the TSCC dataset further highlights its robustness, achieving a recall of 0.9895 with minimal additional training data from TSCC. These results emphasize the potential of fine-tuned GPT-4o-mini as an accurate and cost-effective tool for PII detection in educational data. It offers robust privacy protection while preserving the data's utility for research and pedagogical analysis. Our code is available on GitHub: https://github.com/AnonJD/PrivacyAI

Jie Cao, Chloe Qianhui Zhao, Christian Schunn et al.

Feedback is vital for learning, yet its effectiveness depends not only on its content but also on how it engages students in the learning process. Large Language Models (LLMs) offer novel opportunities to efficiently generate rich, formative feedback, ranging from direct explanations to incrementally layered scaffolding designed to foster learner autonomy. Despite these affordances, it remains unclear whether layered feedback (which sequences encouragement and prompts prior to revealing the correct answer) actually improves engagement and learning outcomes. To address this, we randomly assigned 199 participants to receive either layered or non-layered LLM-generated feedback. We assessed its impact on learning performance, behavioral and cognitive engagement, and affective perceptions, to determine how these factors mediate learning performance. Results indicate that layered feedback elicited slightly higher behavioral engagement and, as anticipated, was perceived as more encouraging and supportive of independence. However, it concurrently induced greater mental effort. Mediation analyses revealed a positive affective pathway driven by perceived encouragement, which was counteracted by a negative behavioral pathway linked to the average number of tasks requiring $\geq 3$ submissions; the cognitive pathway (mental effort) was non-significant. Taken together, layered feedback resulted in significantly poorer learning outcomes compared to non-layered feedback. These findings illuminate a critical trade-off: while layered scaffolding enhances engagement and positive perceptions, it can detrimentally impact actual learning performance. This study contributes nuanced insights for the design of automated, LLM-driven feedback systems by integrating outcome, perception, and mechanism-level analyses.

Ashish Gurung, Ge Gao, Jordan Gutterman et al.

Hybrid human-AI tutoring, where technology and humans jointly facilitate student learning, can be more beneficial than AI-only tutoring. However, preliminary evidence suggests that lower-performing students derive greater benefit from human-AI tutoring than higher-performing students. As such, this study evaluates whether a differentiated tutoring policy can effectively support both groups: human tutors initiate support for lower-performing students, while higher-performing students receive reactive, on-demand support. Using their within-grade median state test scores, we assigned 635 students (grades 5-8) to receive proactive (< median) or reactive ($\geq$ median) tutoring. Using a DiDC design, we compare outcomes across two time periods: fall (AI-only tutoring) and spring (proactive-reactive human-AI tutoring). This quasi-experimental design isolates the effects of proactive-reactive tutoring approaches by comparing the discontinuity in spring outcomes to the fall, where no such discontinuity existed. Using data around the cutoff (Imbens-Kalyanaraman criterion), we find significant overall improvements from human-AI tutoring compared to AI-only baseline: 25% increase in time on task, 36% in skill proficiency, and 61% in academic growth (standardized MAP test). Between proactive and reactive tutoring, we find comparable improvements in time-on-task and skill proficiency. However, proactive tutoring, on average, showed marginally higher MAP growth (75%, p = .065) than reactive tutoring, i.e., proactive tutoring was more beneficial to students farther below the cutoff and helped narrow achievement gaps. Our findings provide evidence that differentiated human-AI tutoring addresses the needs of both groups, offering a practical and cost-effective strategy for scaling hybrid instruction.

Christopher J. MacLellan, Erik Harpstead, Vincent Aleven et al. · gatech

The literature on concept formation has demonstrated that humans are capable of learning concepts incrementally, with a variety of attribute types, and in both supervised and unsupervised settings. Many models of concept formation focus on a subset of these characteristics, but none account for all of them. In this paper, we present TRESTLE, an incremental account of probabilistic concept formation in structured domains that unifies prior concept learning models. TRESTLE works by creating a hierarchical categorization tree that can be used to predict missing attribute values and cluster sets of examples into conceptually meaningful groups. It updates its knowledge by partially matching novel structures and sorting them into its categorization tree. Finally, the system supports mixed-data representations, including nominal, numeric, relational, and component attributes. We evaluate TRESTLE's performance on a supervised learning task and an unsupervised clustering task. For both tasks, we compare it to a nonincremental model and to human participants. We find that this new categorization model is competitive with the nonincremental approach and more closely approximates human behavior on both tasks. These results serve as an initial demonstration of TRESTLE's capabilities and show that, by taking key characteristics of human learning into account, it can better model behavior than approaches that ignore them.

Jionghao Lin, Zifei Han, Danielle R. Thomas et al. · cmu

One-on-one tutoring is widely acknowledged as an effective instructional method, conditioned on qualified tutors. However, the high demand for qualified tutors remains a challenge, often necessitating the training of novice tutors (i.e., trainees) to ensure effective tutoring. Research suggests that providing timely explanatory feedback can facilitate the training process for trainees. However, it presents challenges due to the time-consuming nature of assessing trainee performance by human experts. Inspired by the recent advancements of large language models (LLMs), our study employed the GPT-4 model to build an explanatory feedback system. This system identifies trainees' responses in binary form (i.e., correct/incorrect) and automatically provides template-based feedback with responses appropriately rephrased by the GPT-4 model. We conducted our study on 410 responses from trainees across three training lessons: Giving Effective Praise, Reacting to Errors, and Determining What Students Know. Our findings indicate that: 1) using a few-shot approach, the GPT-4 model effectively identifies correct/incorrect trainees' responses from three training lessons with an average F1 score of 0.84 and an AUC score of 0.85; and 2) using the few-shot approach, the GPT-4 model adeptly rephrases incorrect trainees' responses into desired responses, achieving performance comparable to that of human experts.

Jionghao Lin, Eason Chen, Zeifei Han et al. · cmu

Automated explanatory feedback systems play a crucial role in facilitating learning for a large cohort of learners by offering feedback that incorporates explanations, significantly enhancing the learning process. However, delivering such explanatory feedback in real-time poses challenges, particularly when high classification accuracy for domain-specific, nuanced responses is essential. Our study leverages the capabilities of large language models, specifically Generative Pre-Trained Transformers (GPT), to explore a sequence labeling approach focused on identifying components of desired and less desired praise for providing explanatory feedback within a tutor training dataset. Our aim is to equip tutors with actionable, explanatory feedback during online training lessons. To investigate the potential of GPT models for providing the explanatory feedback, we employed two commonly-used approaches: prompting and fine-tuning. To quantify the quality of highlighted praise components identified by GPT models, we introduced a Modified Intersection over Union (M-IoU) score. Our findings demonstrate that: (1) the M-IoU score effectively correlates with human judgment in evaluating sequence quality; (2) using two-shot prompting on GPT-3.5 resulted in decent performance in recognizing effort-based (M-IoU of 0.46) and outcome-based praise (M-IoU of 0.68); and (3) our optimally fine-tuned GPT-3.5 model achieved M-IoU scores of 0.64 for effort-based praise and 0.84 for outcome-based praise, aligning with the satisfaction levels evaluated by human coders. Our results show promise for using GPT models to provide feedback that focuses on specific elements in their open-ended responses that are desirable or could use improvement.

Zifei FeiFei Han, Jionghao Lin, Ashish Gurung et al. · cmu

One-on-one tutoring is an effective instructional method for enhancing learning, yet its efficacy hinges on tutor competencies. Novice math tutors often prioritize content-specific guidance, neglecting aspects such as social-emotional learning. Social-emotional learning promotes equity and inclusion and nurturing relationships with students, which is crucial for holistic student development. Assessing the competencies of tutors accurately and efficiently can drive the development of tailored tutor training programs. However, evaluating novice tutor ability during real-time tutoring remains challenging as it typically requires experts-in-the-loop. To address this challenge, this preliminary study aims to harness Generative Pre-trained Transformers (GPT), such as GPT-3.5 and GPT-4 models, to automatically assess tutors' ability of using social-emotional tutoring strategies. Moreover, this study also reports on the financial dimensions and considerations of employing these models in real-time and at scale for automated assessment. The current study examined four prompting strategies: two basic Zero-shot prompt strategies, Tree of Thought prompt, and Retrieval-Augmented Generator (RAG) based prompt. The results indicate that the RAG prompt demonstrated more accurate performance (assessed by the level of hallucination and correctness in the generated assessment texts) and lower financial costs than the other strategies evaluated. These findings inform the development of personalized tutor training interventions to enhance the the educational effectiveness of tutored learning.

Sanjit Kakarla, Danielle Thomas, Jionghao Lin et al. · cmu

Research suggests that tutors should adopt a strategic approach when addressing math errors made by low-efficacy students. Rather than drawing direct attention to the error, tutors should guide the students to identify and correct their mistakes on their own. While tutor lessons have introduced this pedagogical skill, human evaluation of tutors applying this strategy is arduous and time-consuming. Large language models (LLMs) show promise in providing real-time assessment to tutors during their actual tutoring sessions, yet little is known regarding their accuracy in this context. In this study, we investigate the capacity of generative AI to evaluate real-life tutors' performance in responding to students making math errors. By analyzing 50 real-life tutoring dialogues, we find both GPT-3.5-Turbo and GPT-4 demonstrate proficiency in assessing the criteria related to reacting to students making errors. However, both models exhibit limitations in recognizing instances where the student made an error. Notably, GPT-4 tends to overidentify instances of students making errors, often attributing student uncertainty or inferring potential errors where human evaluators did not. Future work will focus on enhancing generalizability by assessing a larger dataset of dialogues and evaluating learning transfer. Specifically, we will analyze the performance of tutors in real-life scenarios when responding to students' math errors before and after lesson completion on this crucial tutoring skill.

Danielle R. Thomas, Conrad Borchers, Sanjit Kakarla et al. · cmu

The role of multiple-choice questions (MCQs) as effective learning tools has been debated in past research. While MCQs are widely used due to their ease in grading, open response questions are increasingly used for instruction, given advances in large language models (LLMs) for automated grading. This study evaluates MCQs effectiveness relative to open-response questions, both individually and in combination, on learning. These activities are embedded within six tutor lessons on advocacy. Using a posttest-only randomized control design, we compare the performance of 234 tutors (790 lesson completions) across three conditions: MCQ only, open response only, and a combination of both. We find no significant learning differences across conditions at posttest, but tutors in the MCQ condition took significantly less time to complete instruction. These findings suggest that MCQs are as effective, and more efficient, than open response tasks for learning when practice time is limited. To further enhance efficiency, we autograded open responses using GPT-4o and GPT-4-turbo. GPT models demonstrate proficiency for purposes of low-stakes assessment, though further research is needed for broader use. This study contributes a dataset of lesson log data, human annotation rubrics, and LLM prompts to promote transparency and reproducibility.

Danielle R. Thomas, Conrad Borchers, Sanjit Kakarla et al. · cmu

Equity is a core concern of learning analytics. However, applications that teach and assess equity skills, particularly at scale are lacking, often due to barriers in evaluating language. Advances in generative AI via large language models (LLMs) are being used in a wide range of applications, with this present work assessing its use in the equity domain. We evaluate tutor performance within an online lesson on enhancing tutors' skills when responding to students in potentially inequitable situations. We apply a mixed-method approach to analyze the performance of 81 undergraduate remote tutors. We find marginally significant learning gains with increases in tutors' self-reported confidence in their knowledge in responding to middle school students experiencing possible inequities from pretest to posttest. Both GPT-4o and GPT-4-turbo demonstrate proficiency in assessing tutors ability to predict and explain the best approach. Balancing performance, efficiency, and cost, we determine that few-shot learning using GPT-4o is the preferred model. This work makes available a dataset of lesson log data, tutor responses, rubrics for human annotation, and generative AI prompts. Future work involves leveling the difficulty among scenarios and enhancing LLM prompts for large-scale grading and assessment.

Conrad Borchers, Danielle R. Thomas, Jionghao Lin et al. · cmu

Large Language Models (LLMs) like GPT-4o can help automate text classification tasks at low cost and scale. However, there are major concerns about the validity and reliability of LLM outputs. By contrast, human coding is generally more reliable but expensive to procure at scale. In this study, we propose a hybrid solution to leverage the strengths of both. We combine human-coded data and synthetic LLM-produced data to fine-tune a classical machine learning classifier, distilling both into a smaller BERT model. We evaluate our method on a human-coded test set as a validity measure for LLM output quality. In three experiments, we systematically vary LLM-generated samples' size, variety, and consistency, informed by best practices in LLM tuning. Our findings indicate that augmenting datasets with synthetic samples improves classifier performance, with optimal results achieved at an 80% synthetic to 20% human-coded data ratio. Lower temperature settings of 0.3, corresponding to less variability in LLM generations, produced more stable improvements but also limited model learning from augmented samples. In contrast, higher temperature settings (0.7 and above) introduced greater variability in performance estimates and, at times, lower performance. Hence, LLMs may produce more uniform output that classifiers overfit to earlier or produce more diverse output that runs the risk of deteriorating model performance through information irrelevant to the prediction task. Filtering out inconsistent synthetic samples did not enhance performance. We conclude that integrating human and LLM-generated data to improve text classification models in assessment offers a scalable solution that leverages both the accuracy of human coding and the variety of LLM outputs.

Danielle R. Thomas, Conrad Borchers, Jionghao Lin et al. · cmu

Tutoring improves student achievement, but identifying and studying what tutoring actions are most associated with student learning at scale based on audio transcriptions is an open research problem. This present study investigates the feasibility and scalability of using generative AI to identify and evaluate specific tutor moves in real-life math tutoring. We analyze 50 randomly selected transcripts of college-student remote tutors assisting middle school students in mathematics. Using GPT-4, GPT-4o, GPT-4-turbo, Gemini-1.5-pro, and LearnLM, we assess tutors' application of two tutor skills: delivering effective praise and responding to student math errors. All models reliably detected relevant situations, for example, tutors providing praise to students (94-98% accuracy) and a student making a math error (82-88% accuracy) and effectively evaluated the tutors' adherence to tutoring best practices, aligning closely with human judgments (83-89% and 73-77%, respectively). We propose a cost-effective prompting strategy and discuss practical implications for using large language models to support scalable assessment in authentic settings. This work further contributes LLM prompts to support reproducibility and research in AI-supported learning.

Danielle R. Thomas, Conrad Borchers, Shambhavi Bhushan et al.

Large language models (LLMs) are increasingly used to generate feedback, yet their impact on learning remains underexplored, especially compared to existing feedback methods. This study investigates how on-demand LLM-generated explanatory feedback influences learning in seven scenario-based tutor training lessons. Analyzing over 2,600 lesson completions from 885 tutor learners, we compare posttest performance among learners across three groups: learners who received feedback generated by gpt-3.5-turbo, those who declined it, and those without access. All groups received non-LLM corrective feedback. To address potential selection bias-where higher-performing learners may be more inclined to use LLM feedback-we applied propensity scoring. Learners with a higher predicted likelihood of engaging with LLM feedback scored significantly higher at posttest than those with lower propensity. After adjusting for this effect, two out of seven lessons showed statistically significant learning benefits from LLM feedback with standardized effect sizes of 0.28 and 0.33. These moderate effects suggest that the effectiveness of LLM feedback depends on the learners' tendency to seek support. Importantly, LLM feedback did not significantly increase completion time, and learners overwhelmingly rated it as helpful. These findings highlight LLM feedback's potential as a low-cost and scalable way to improve learning on open-ended tasks, particularly in existing systems already providing feedback without LLMs. This work contributes open datasets, LLM prompts, and rubrics to support reproducibility.

Eason Chen, Chuangji Li, Shizhuo Li et al. · cmu

Technology-enhanced learning environments often help students retrieve relevant learning content for questions arising during self-paced study. Large language models (LLMs) have emerged as novel aids for information retrieval during learning. While LLMs are effective for general-purpose question-answering, they typically lack alignment with the domain knowledge of specific course materials such as textbooks and slides. We investigate Retrieval-Augmented Generation (RAG) and GraphRAG, a knowledge graph-enhanced RAG approach, for page-level question answering in an undergraduate mathematics textbook. While RAG has been effective for retrieving discrete, contextually relevant passages, GraphRAG may excel in modeling interconnected concepts and hierarchical knowledge structures. We curate a dataset of 477 question-answer pairs, each tied to a distinct textbook page. We then compare the standard embedding-based RAG methods to GraphRAG for evaluating both retrieval accuracy-whether the correct page is retrieved-and generated answer quality via F1 scores. Our findings show that embedding-based RAG achieves higher retrieval accuracy and better F1 scores compared to GraphRAG, which tends to retrieve excessive and sometimes irrelevant content due to its entity-based structure. We also explored re-ranking the retrieved pages with LLM and observed mixed results, including performance drop and hallucinations when dealing with larger context windows. Overall, this study highlights both the promises and challenges of page-level retrieval systems in educational contexts, emphasizing the need for more refined retrieval methods to build reliable AI tutoring solutions in providing reference page numbers.

Ruiwei Xiao, Xinying Hou, Ying-Jui Tseng et al.

As Artificial Intelligence (AI) becomes increasingly integrated into daily life, there is a growing need to equip the next generation with the ability to apply, interact with, evaluate, and collaborate with AI systems responsibly. Prior research highlights the urgent demand from K-12 educators to teach students the ethical and effective use of AI for learning. To address this need, we designed an Large-Language Model (LLM)-based module to teach prompting literacy. This includes scenario-based deliberate practice activities with direct interaction with intelligent LLM agents, aiming to foster secondary school students' responsible engagement with AI chatbots. We conducted two iterations of classroom deployment in 11 authentic secondary education classrooms, and evaluated 1) AI-based auto-grader's capability; 2) students' prompting performance and confidence changes towards using AI for learning; and 3) the quality of learning and assessment materials. Results indicated that the AI-based auto-grader could grade student-written prompts with satisfactory quality. In addition, the instructional materials supported students in improving their prompting skills through practice and led to positive shifts in their perceptions of using AI for learning. Furthermore, data from Study 1 informed assessment revisions in Study 2. Analyses of item difficulty and discrimination in Study 2 showed that True/False and open-ended questions could measure prompting literacy more effectively than multiple-choice questions for our target learners. These promising outcomes highlight the potential for broader deployment and highlight the need for broader studies to assess learning effectiveness and assessment design.

Danielle R. Thomas, Conrad Borchers, Kenneth R. Koedinger

Humans can be notoriously imperfect evaluators. They are often biased, unreliable, and unfit to define "ground truth." Yet, given the surging need to produce large amounts of training data in educational applications using AI, traditional inter-rater reliability (IRR) metrics like Cohen's kappa remain central to validating labeled data. IRR remains a cornerstone of many machine learning pipelines for educational data. Take, for example, the classification of tutors' moves in dialogues or labeling open responses in machine-graded assessments. This position paper argues that overreliance on human IRR as a gatekeeper for annotation quality hampers progress in classifying data in ways that are valid and predictive in relation to improving learning. To address this issue, we highlight five examples of complementary evaluation methods, such as multi-label annotation schemes, expert-based approaches, and close-the-loop validity. We argue that these approaches are in a better position to produce training data and subsequent models that produce improved student learning and more actionable insights than IRR approaches alone. We also emphasize the importance of external validity, for example, by establishing a procedure of validating tutor moves and demonstrating that it works across many categories of tutor actions (e.g., providing hints). We call on the field to rethink annotation quality and ground truth--prioritizing validity and educational impact over consensus alone.

Jionghao Lin, Jiarui Rao, Yiyang Zhao et al. · cmu

We explore the automatic generation of interactive, scenario-based lessons designed to train novice human tutors who teach middle school mathematics online. Employing prompt engineering through a Retrieval-Augmented Generation approach with GPT-4o, we developed a system capable of creating structured tutor training lessons. Our study generated lessons in English for three key topics: Encouraging Students' Independence, Encouraging Help-Seeking Behavior, and Turning on Cameras, using a task decomposition prompting strategy that breaks lesson generation into sub-tasks. The generated lessons were evaluated by two human evaluators, who provided both quantitative and qualitative evaluations using a comprehensive rubric informed by lesson design research. Results demonstrate that the task decomposition strategy led to higher-rated lessons compared to single-step generation. Human evaluators identified several strengths in the LLM-generated lessons, including well-structured content and time-saving potential, while also noting limitations such as generic feedback and a lack of clarity in some instructional sections. These findings underscore the potential of hybrid human-AI approaches for generating effective lessons in tutor training.

Megan Gu, Chloe Qianhui Zhao, Claire Liu et al. · cmu

Our study introduces an automated system leveraging large language models (LLMs) to assess the effectiveness of five key tutoring strategies: 1. giving effective praise, 2. reacting to errors, 3. determining what students know, 4. helping students manage inequity, and 5. responding to negative self-talk. Using a public dataset from the Teacher-Student Chatroom Corpus, our system classifies each tutoring strategy as either being employed as desired or undesired. Our study utilizes GPT-3.5 with few-shot prompting to assess the use of these strategies and analyze tutoring dialogues. The results show that for the five tutoring strategies, True Negative Rates (TNR) range from 0.655 to 0.738, and Recall ranges from 0.327 to 0.432, indicating that the model is effective at excluding incorrect classifications but struggles to consistently identify the correct strategy. The strategy \textit{helping students manage inequity} showed the highest performance with a TNR of 0.738 and Recall of 0.432. The study highlights the potential of LLMs in tutoring strategy analysis and outlines directions for future improvements, including incorporating more advanced models for more nuanced feedback.