Many students disengage from mathematics throughout their school careers, a pattern often attributed to declines in motivation. While long-term motivational trajectories are well documented, far less is known about short-term motivational fluctuations in self-paced online courses; a setting that becomes increasingly popular in mathematics education and which places high demands on learners’ autonomy, possibly increasing the risk of disengagement when motivation decreases. Drawing on situated expectancy-value theory, we investigated the relations between dispositional (trait-like) motivation, situational (state-like) motivation, and parental help in an extracurricular enrichment online course designed to foster children’s proof competency. Longitudinal data from 159 talented primary school students indicated considerable within-person variability in motivation. Dispositional motivation at the beginning of the course was only modestly related to children’s average situational motivation during the course. Notably, across different course sections, motivational declines seemed more pronounced when students encountered cognitively demanding content (e.g., construction of formal-deductive proofs). For situational utility value—children’s perceived usefulness of the specific course section for school life—parental help emerged as a critical moderator, significantly buffering declines between two mid-course sections. Critically, our data indicated high dropout rates among children. Survival analyses suggested that greater situational motivation (particularly self-concept), but not greater dispositional motivation, reduced the risk of dropout, highlighting the importance of momentary motivational states in sustaining engagement.

Generative artificial intelligence (genAI) systems are increasingly integral to epistemic processes such as hypothesis generation, explanation construction, and decision-making. Although they reliably enhance performance, emerging evidence reveals a metacognitive dilemma: as external generative capacity increases, internal monitoring, calibration, and cognitive engagement may decline. This reflects a redistribution of cognitive control within distributed human-AI systems that cannot be explained by automation bias or reliance on algorithms alone. We propose the AIRIS (AI-Augmented Inquiry and Regulation in Hybrid Systems) framework to analyze this dilemma and specify where regulatory intervention can counteract it. AIRIS is a multi-level control allocation architecture specifying the conditions under which epistemic agency can be preserved in hybrid generative systems. Drawing on distributed cognition, cognitive load theory, multimedia learning, and self-regulated learning, it identifies seven interacting mechanisms through which hybrid cognition may become destabilized, from delegation and calibration drift to motivational-affective drift. Five regulatory operators (Anticipate, Interrogate, Reflect, Integrate, and Synthesize) target internal generative engagement at points of emerging instability. The architecture does not itself improve learning; it specifies what must remain in place for genAI-supported work to sustain understanding, whether through instructional design, teacher guidance, or learners' own regulation. We derive testable propositions concerning the seven mechanisms and the five operators, reframing AI augmentation as a problem of control allocation in distributed generative systems. Beyond theory, AIRIS offers a research agenda, a design framework for genAI-integrated learning environments, and a conceptual toolkit for the governance of hybrid human-AI cognition.

Artificial intelligence (AI) is transforming classrooms, and promising adaptive teaching and personalized education. However, AI tools may also weaken students’ ability to think and learn independently. Schools and universities should use AI as a coach to support self-regulated learning.

Why do some self-regulated learning (SRL) interventions seem to benefit less competent students more than their competent peers (i.e., compensatory effect), but others seem to benefit only the already competent students (i.e., Matthew effects)? We propose the Resource-Intervention Match (RIM) framework to explain these differential outcomes. Intervention effects depend on the (mis-)match between learners' existing SRL resources and specific intervention features. We conceptualize SRL resources as comprising three components: metacognitive knowledge, metacognitive skills, and motivational-affective resources. When learners' resources align with intervention demands, learners experience gains in performance; misalignment creates non-productive experiences that hinder progress. A critical but overlooked factor is metacognitive experiences (e.g., feelings of difficulty, confidence, and satisfaction) that emerge during learning. These experiences serve as the mediating mechanism through which resource-intervention (mis-)matches influence intervention outcomes. The RIM framework provides researchers and practitioners with a systematic approach to diagnosing, predicting, and optimizing SRL intervention effects across individual differences. Educational relevance and implications statement This research explains why some learning interventions help struggling students catch up (compensatory effects) whereas others primarily benefit already-successful students (Matthew effects). We found that effectiveness depends on matching support to specific gaps in students' self-regulated learning: their knowledge about effective strategies, their ability to actually use these strategies, and their motivation to persist through challenges. Teachers can assess these three components separately through questionnaires and classroom observation, then provide personalized support that adjusts based on each student's needs and gradually fades as they develop skills. This approach transforms students from those requiring constant external guidance into independent learners who can systematically figure out which study approaches work best for their individual needs.

IntroductionThe purpose of this study is to explore the effects of a randomized control trial designed to test the effect of a brief intervention used to improve self-regulated learning (SRL) in gateway biology courses using joint estimation of graphical models.MethodsStudents (N = 265; n = 136) from three sections of a hybrid-format introductory biology course were randomly assigned to participate in the multimedia science of learning to learn or a multimedia control condition. All participants completed a self-report battery of motivational measures. Course performance data was also collected.ResultsNetwork structures of motivation variables were estimated in two sub-groups (Treatment and Control). These networks showed a high level of correspondence in the relative magnitudes of the edge weights, however there were non-trivial differences in the edge weights between groups that may be attributed to the treatment and differences in predictability. While these findings suggest meaningful differences in motivational structures, the relatively small sample size may limit the stability of the estimated network models. The SRL strategy based interventions may have positioned the students motivationally to approach the challenging exam through activating the role of value and self-efficacy in their learning.DiscussionMany of the ways analyses of typical intervention studies are conducted ignore the underlying complexity of what motivates individuals. This study provides preliminary evidence how Gaussian Graphical Modeling may be valuable in preserving the integrity of complex systems and examining relevant shifts in variations between motivational systems between groups and individuals.

Despite extensive efforts to reduce STEM (science, technology, engineering, and mathematics) dropout, the United States still faces a shortage of STEM professionals. Prior research has mainly focused on long-term academic trajectories, but less attention has been given to challenging introductory STEM courses known to prevent many students from pursuing STEM degrees. Existing STEM dropout prediction models primarily rely on demographics and prior academic performance, neglecting the role of motivation in dropout decisions. We address these gaps by developing a prediction model that integrates prior academic performance, motivation, and early course performance in a course at a community college with an academically underprepared and ethnically diverse student population. Results show that dropout patterns do not vary significantly by gender, ethnicity, or first-generation status. However, students' perceived cost of course engagement interacts with early course performance to predict dropout intentions. We argue for incorporating motivational factors into dropout models and offer recommendations for prediction modeling and intervention.

Background: Students in upper primary grades must move beyond basic comprehension toward high-level comprehension (HLC) of text as they read. Small-group, text-based discussions provide opportunities for students to develop their critical analytic thinking and argumentation, supporting their HLC. Aims: We explored the extent to which groups of upper primary students evidenced growth on indicators of HLC as they engaged in small-group, text-based discussions over a school year, while also examining grade-level and text genre differences. Sample: Participants included fourth-(n = 64) and fifth-grade (n = 69) students. Methods: We employed a single-group, longitudinal design, whereby Quality Talk was embedded into the language arts curriculum of six upper elementary classrooms. Video-recorded discussions (n = 371) were transcribed. We employed an artificial intelligence (AI) powered coding approach to identify indicators of HLC in the discussion transcripts. Results: Groups of upper primary students, on average, evidenced growth in the rates of HLC indicators over the school year. Groups composed of fifth-grade students, on average, had higher elaborated explanation rates than fourth-grade students, and all students, on average, produced a higher rate of elaborated explanations for discussions based on mixed genre versus expository genre texts. Conclusions: Findings from this study contribute to a growing body of literature about grade-level differences in upper primary grades, as well as the influence of text genre on indicators of HLC present within small-group discussions. Notably, the study also employed a novel, AI-powered coding approach for our discourse analysis, which warrants further exploration in future research.

Barzilai and Chinn (2018) have advocated for framing education more overtly around epistemic matters to help students learn how to create worthwhile epistemic aims, which they achieve by using appropriate epistemic standards and enacting suitable reliable processes. It is not clear, though, whether existing disciplinary education already addresses aspects of epistemic education. The current study examined associations between prior students' disciplinary achievement and use of epistemic ideals and reliable processes when judging the quality of published research, and planning their own research study. Seventy-one psychology majors completed an Epistemic Criteria and Strategies Assessment, which elicited their abilities to judge the quality of extant published research, and to use their knowledge of research design to design a study of their own within the discipline of psychology. Responses were coded for students' application of epistemic ideals and reliable processes for producing justifiable knowledge. Results indicated that students did engage in epistemic cognition. In doing so, there were positive associations between prior achievement in psychology courses and more frequent employment of epistemic ideals to evaluate the quality of published research, and with greater evidence for students' metacognitive reflection on their epistemic ideals and reliable processes when designing a study in psychology. These findings illustrate the importance of incorporating epistemic education in curricula, with implications for future epistemic cognition research.

The increasing use of learning management systems (LMSs) generates vast amounts of clickstream data, opening new avenues for predicting learner performance. Traditionally, LMS predictive analytics have relied on either supervised machine learning or Markov models to classify learners based on predicted learning outcomes. Machine learning excels at pattern recognition but often overlooks temporal learning dynamics and obscures the reasoning behind predictions due to the black-box nature of many algorithms. Alternatively, Markov models provide an effective solution by capturing temporal learning dynamics for prediction, uncovering distinctive learning patterns between high and low performers. Despite these advantages, Markov model classification struggles with the heterogeneity of learning sequences, limiting its broad applicability. To address these limitations and bridge the gap between the two dominant approaches, we propose a hybrid framework: sequence-based Markov machine learning classification (seqMAC). Leveraging early-stage clickstream data, seqMAC provides an interpretable sequence classification method that captures critical behavioural transitions and identifies distinct learning patterns across performance groups. Tested on six LMS samples, seqMAC effectively identified at-risk students despite sequence heterogeneity, uncovering key predictive learning dynamics that differentiate performance groups. It also demonstrated promising generalizability, accurately identifying future at-risk students based on historical clickstream data.

The increasing use of learning management systems (LMSs) generates vast amounts of clickstream data, opening new avenues for predicting learner performance. Traditionally, LMS predictive analytics have relied on either supervised machine learning or Markov models to classify learners based on predicted learning outcomes. Machine learning excels at pattern recognition but often overlooks temporal learning dynamics and obscures the reasoning behind predictions due to the black-box nature of many algorithms. Alternatively, Markov models provide an effective solution by capturing temporal learning dynamics for prediction, uncovering distinctive learning patterns between high and low performers. Despite these advantages, Markov model classification struggles with the heterogeneity of learning sequences, limiting its broad applicability. To address these limitations and bridge the gap between the two dominant approaches, we propose a hybrid framework: sequence-based Markov machine learning classification (seqMAC). Leveraging early-stage clickstream data, seqMAC provides an interpretable sequence classification method that captures critical behavioural transitions and identifies distinct learning patterns across performance groups. Tested on six LMS samples, seqMAC effectively identified at-risk students despite sequence heterogeneity, uncovering key predictive learning dynamics that differentiate performance groups. It also demonstrated promising generalizability, accurately identifying future at-risk students based on historical clickstream data.

The shift towards active pedagogies in higher education that emphasize students’ engagement in their own learning in and outside of the classroom has increased the ubiquity of online learning and assessment platforms for engaging students in such learning. Online learning requires self-regulated learning, which is a cyclical and temporal process in which students plan, monitor, and control their cognition, motivation, behavior, and affect in pursuit of their learning goals. Help-seeking is a particularly important regulation strategy when learning online, but few researchers have examined the cyclical and temporal nature of help-seeking processes when students learn in an online learning and assessment platform. We conducted a micro-level analysis of the temporal help-seeking behaviors of 488 undergraduates in an online learning and assessment platform to explore how they sought help during learning and identify those who struggled in such a context. This exploratory study includes two levels of analysis, frequency analysis and process mining, to triangulate patterns of help-seeking transitions observed in an online learning and assessment platform and relate those to learning. Results indicated (1) less successful learners demonstrated an increase in incorrect submissions and transitions to and from incorrect submissions, and (2) lower performers used more maladaptive help-seeking strategies during independent learning before classes (e.g., via repetitive use of solutions). The findings demonstrate the benefits of applying multiple learning analytics methods to inform robust interpretations of micro-level self-regulated learning and suggest that such modeling can help prepare interventions that support undergraduate students’ effective use of help-seeking when learning online.

Educators, families, and students continue to debate whether homework promotes academic achievement. A resolution to this debate has proven elusive, given the often-mixed findings of the relationship between homework behavior, typically measured with often-unreliable student self-reports and achievement. We argue better estimates of these relationships require (a) changes to what data are collected to measure homework behavior and (b) more theory-informed ways to model those data. Thus, in this article, we pursued what Marsh and Hau (2007) called substantive-methodological synergy. We grounded our substantive investigation in Trautwein et al.'s (2006) Homework Model, wherein student characteristics and motivation predict homework behaviors (i.e., homework effort, homework time), which in turn predict achievement. To better understand students' homework behavior, we used digital tools that produced trace data that could be understood and modeled via theory-informed learning analytics. We collected homework behavior data and subsequent achievements from 507 German academic-track school students who used an intelligent tutoring system to learn English as a foreign language. Our initial analyses showed that theory-aligned digital trace data captured unique information beyond self-report data. Then, we found homework effort, as conceptualized in the Homework Model and captured via theory-informed learning analytics, predicted academic performance, whereas homework time did not. Overall, behavioral trace measures of homework effort were more predictive than self-reports. These findings help to clarify the mixed findings in the homework literature and illustrate the benefits of substantive-methodological synergy between theory and learning analytic methods.