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 (≥ 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

This paper introduces a scalable causal inference framework for estimating the immediate, session-level effects of on-demand human tutoring embedded within adaptive learning systems. Because students seek assistance at moments of difficulty, conventional evaluation is confounded by self-selection and time-varying knowledge states. We address these challenges by integrating principled analytic sample construction with Deep Knowledge Tracing (DKT) to estimate latent mastery, followed by doubly robust estimation using Causal Forests. Applying this framework to over 5,000 middle-school mathematics tutoring sessions, we find that requesting human tutoring increases next-problem correctness by approximately 4 percentage points and accuracy on the subsequent skill encountered by approximately 3 percentage points, suggesting that the effects of tutoring have proximal transfer across knowledge components. This effect is robust to various forms of model specification and potential unmeasured confounders. Notably, these effects exhibit significant heterogeneity across sessions and students, with session-level effect estimates ranging from $-20.25pp$ to $+19.91pp$. Our follow-up analyses suggest that typical behavioral indicators, such as student talk time, do not consistently correlate with high-impact sessions. Furthermore, treatment effects are larger for students with lower prior mastery and slightly smaller for low-SES students. This framework offers a rigorous, practical template for the evaluation and continuous improvement of on-demand human tutoring, with direct applications for emerging AI tutoring systems.

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

Sustained effort is essential for realizing the benefits of intelligent tutoring systems (ITS), yet many learners disengage or underuse available practice time. We introduce engagement forecasting as a supervised prediction task based on ITS logs, targeting two outcomes central to effort and learning progress: minutes practiced per week and new skills mastered per week. Using interaction log data from 425 middle-school students over a school year, we benchmark fifteen predictors including regressions, decision trees, and neural networks. We show that these feature-based models reduce mean absolute error (MAE) by 22-33

Learning analytics can guide human tutors to efficiently address motivational barriers to learning that AI systems struggle to support. Students become more engaged when they receive human attention. However, what occurs during short interventions, and when are they most effective? We align student-tutor dialogue transcripts with MATHia tutoring system log data to study brief human-tutor interactions on Zoom drawn from 2,075 hours of 191 middle school students' classroom math practice. Mixed-effect models reveal that engagement, measured as successful solution steps per minute, is higher during a human-tutor visit and remains elevated afterward. Visit length exhibits diminishing returns: engagement rises during and shortly after visits, irrespective of visit length. Timing also matters: later visits yield larger immediate lifts than earlier ones, though an early visit remains important to counteract engagement decline. We create analytics that identify which tutor-student dialogues raise engagement the most. Qualitative analysis reveals that interactions with concrete, stepwise scaffolding with explicit work organization elevate engagement most strongly. We discuss implications for resource-constrained tutoring, prioritizing several brief, well-timed check-ins by a human tutor while ensuring at least one early contact. Our analytics can guide the prioritization of students for support and surface effective tutor moves in real-time.

High-quality tutoring is among the most impactful instructional interventions in education. However, these programs remain difficult to scale effectively, and the specific “moves” underlying quality tutoring are understudied due to historical data scarcity. Despite extensive research, progress is hindered by challenges in data de-identification, multimodal analysis, and the predictive modeling of student outcomes. The emergence of Artificial Intelligence is fundamentally shifting the capacity to scale and study tutoring, offering transformative potential alongside significant pitfalls. This workshop, led by the National Tutoring Observatory, the SCALE Initiative, and the LEVI HAT project, brings together researchers, providers, and practitioners to explore human and AI tutoring systems. Featuring sessions on open-source data, infrastructure benchmarks, and synthetic students, the workshop aims to foster collaboration that ensures the future of instruction is grounded in rigorous empirical science.

There exist numerous tutor training platforms. However, few provide AI-driven training and evaluation for human tutors based on real-life performance. We present an AI-driven system that assesses both open responses during training and authentic real-life tutoring. Unlike platforms that only assess learning through online training or simulations, our system utilizes Generative AI (Gemini-2.5-pro) to analyze transcriptions of authentic tutoring, measuring the transfer of tutor skills to real-life application. Human tutors instructing students remotely in math (N=86) completed six scenario-based lessons, averaging a significant 7.4

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 ( κ ), 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., κ > 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.

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 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 TalkMoves achieved higher overall inter-rater reliability than the AI-human codebook ( κ = 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 non-verbal 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.

Assessment is foundational to learning analytics, especially in evaluating instructional interventions and guiding improvement in online learning environments. With the growing use of large language models (LLMs) to score open-ended responses, questions arise about the reliability of these model-generated scores, particularly in short pre-post formats where learners are expected to improve. This study introduces a novel method for estimating test reliability that adjusts for learning gains using a Rasch-based split-half approach. We validated this approach through simulation under realistic conditions of missing data and score change, showing tangible improvements in reliability estimation compared to baseline methods. Applying this method to a dataset of 985 tutors completing 12 online lessons, we find that GPT-4-based scoring achieves satisfactory reliability, with open-ended responses (0.733) outperforming multiple-choice items (0.652). Both item types jointly yielded the highest reliability (0.774). Hence, as few as 14 open-ended items (across an average of 3-4 completed lessons) were sufficient to surpass common reliability thresholds of 0.7 or higher. Principal component analysis revealed a skill structure with a strong primary dimension shared across almost all lessons and interpretable subdimensions—socio-emotional, cognitive, and fairness-related tutoring skills—supporting a bifactor-like model. These findings demonstrate that GPT-4 and similar LLMs can be effectively used for formative assessment of complex instructional skills in online and personalized learning contexts, provided their reliability is empirically verified. This study contributes an open-source, learning-aware framework for scalable and reliable AI-supported assessment in learning analytics contexts.

The increasing availability of large language models (LLMs) has raised concerns about their potential misuse in online learning. While tools for detecting LLM-generated text exist and are widely used by researchers and educators, their reliability varies. Few studies have compared the accuracy of detection methods, defined criteria to identify content generated by LLM, or evaluated the effect on learner performance from LLM misuse within learning. In this study, we define LLM-generated text within open responses as those produced by any LLM without paraphrasing or refinement, as evaluated by human coders. We then fine-tune GPT-4o to detect LLM-generated responses and assess the impact on learning from LLM misuse. We find that our fine-tuned LLM outperforms the existing AI detection tool GPTZero, achieving an accuracy of 80

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.

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.

Automated grading of learners' open responses remains challenging due to the complexity of language and the subjective nature of human evaluation. Recent advances in generative AI, particularly large language models (LLMs), offer new possibilities to improve assessment. Off-the-shelf LLMs, such as GPT-4, have been applied to this task, as well as dedicated education-oriented models, such as LearnLM. However, little is known about their effectiveness compared to general-purpose models. In this study, we evaluate GPT-4o, GPT-4-turbo, and Gemini-Pro and compare their performance to LearnLM to determine their effectiveness in assessing learning, specifically the professional development of adult tutors. We find that LearnLM outperforms other models on tasks requiring tutor learners to predict the most appropriate response to students. We hypothesize that this is due to the model's fine-tuning on tutorstudent interaction data and suggest that LearnLM may be particularly useful in scenario-based tutor training. To further improve automated assessment methods, we challenge the concept of human "ground truth" by proposing alternative validation methods. Specifically, we introduce a predictive validity method by relating open-response scores with corresponding multiple-choice scores that demonstrate statistically significant and moderate correlations, particularly with LearnLM. Our novel method demonstrates predictive validity but should be combined with additional measures to ensure a more comprehensive assessment. This study contributes an open source dataset, human annotation rubrics, and LLM prompts, to improve future assessment applications of LLMs.

Effective teaching is among the most powerful influences on student learning, but scientific progress in understanding effective teaching moves has been held back by insufficient data on teaching. Despite extensive research efforts, progress is hindered by persistent challenges related to data de-identification and preprocessing, annotation and segmentation, multimodal analysis, predictive and causal modeling of student outcomes. Addressing these barriers requires a concerted, interdisciplinary approach. The National Tutoring Observatory (NTO) is a first-of-its-kind research infrastructure designed to unite researchers, developers, tutoring providers, and educational organizations in tackling common barriers to uncovering the dynamics of effective tutoring moves. The NTO is spearheading the creation of the Million Tutor Moves dataset, the largest open-access collection of tutoring interactions, leveraging artificial intelligence to unlock insights that accelerate the science of teaching at scale. This workshop aims to bring together the Learning at Scale community to share progress, identify common challenges, and explore collaborative solutions. The agenda will feature presentations of accepted papers, interactive demos, and a moderated panel bringing together researchers, developers, and tutoring providers. This workshop aims to advance a shared vision for uncovering the fundamental principles of impactful tutoring and teaching through the power of collaborative research and data-driven discovery.

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.

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 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.

High-impact (human) tutoring and computer-based AI tutors are widely recognized for their effectiveness in supporting learning. However, human tutoring is costly and difficult to scale, whereas AI tutors vary widely in their ability to adapt to students' academic and motivational needs. Our study presents a formative evaluation of a year-long implementation of virtual human-AI tutoring during the classroom use of AI tutors. Using year-long log data and standardized state tests, we examine the real-world impact of human-AI tutoring through measures of learning both within the AI tutor and on external standardized assessments. Through propensity score matching, we compare 356 seventh-grade students who received human-AI tutoring with 317 from the previous school year who received AI-only tutoring. The human-AI group demonstrated significantly higher growth and was 0.36 grade levels ahead by year's end. Although there was no overall difference in state test scores, we found a significant interaction between human-AI tutoring and time-on-task (i.e., AI tutor use). For each standard deviation (3.26 h) increase in AI tutor use, the human-AI group improved by 0.28 standard deviations on state tests, compared to 0.06 for the AI-only group. This finding suggests that human tutors enhance the benefits of AI tutors, with gains increasing with time-on-task. Our findings replicate prior studies on human-AI tutoring over a longer time scale, spanning an entire school year. An important insight of this work is that students' AI tutor usage data, particularly time-on-task, can serve as a valuable indicator of learning progress as well as a measure for identifying students in need of additional support to fully benefit from AI tutors.

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.

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.