Sustaining independent student practice in classrooms using adaptive learning software remains a central challenge in realizing the potential of AI-enhanced learning environments. Goal setting, that is, defining clear, quantifiable objectives, can increase student effort. However, goal setting is rarely integrated with the data-driven feedback and orchestration capabilities of modern learning systems. This thesis investigates how goal setting can be extended into adaptive goal support: routines that embed personalized targets, feedback, and adjustment into regular classroom practice to improve student effort and learning.

This dissertation leverages fine-grained platform data to generate weekly practice targets, support brief goal review, and provide adaptive recommendations grounded in prior performance. Empirical studies trace the evolution of this approach from paper-based goal contracts to fully data-driven implementations with automated feedback and classroom-level reward structures independently sustained by teachers.

Across quasi-experimental studies in middle school mathematics, adaptive goal support is robustly associated with increased student practice and skill mastery. In a large-scale Fall 2025 deployment across six middle schools (with four schools including a no-goal baseline), interrupted time series models show substantial positive shifts in weekly minutes practiced and skills completed following implementation. Implementation-intention scaffolding further predicts practice, with effects concentrated among students with lower baseline practice time and attenuated at higher levels of baseline practice time. Baseline intrinsic motivation for mathematics shows a modest positive association with intervention benefit, though motivational measures explain limited variance overall. Gains are also modestly associated with increases in math self-efficacy. Qualitative analysis of students' implementation-intention responses further indicated that students most often aimed to translate goals into environmental regulation strategies involving managing distractions, social regulation, and help-seeking. Finally, students' selected goals closely track system recommendations, explaining 61% of the variance in selected effort goals and 75% of the variance in selected progress goals.

Overall, the findings demonstrate that analytics-informed adaptive goal support is a scalable and effective complement to adaptive middle school mathematics practice software. More broadly, this dissertation argues that educational AI systems should adapt to the behavioral and organizational processes that shape whether sustained practice occurs. The findings further motivate a three-component theoretical model of effort-aware adaptivity in classroom learning environments. Adaptive goal support appears to improve outcomes by (1) increasing the amount of time students are willing to spend practicing, (2) increasing the productivity and focus of engagement during already allocated practice time, and (3) balancing these benefits against the opportunity costs introduced by scaffolding routines themselves within constrained instruction time budgets. Together, the studies position effort regulation as an important and underdeveloped dimension of educational adaptivity and point toward future educational AI systems that personalize support for student persistence.

103 pages

Thesis Committee:
Vincent Aleven (Co-Chair)
Kenneth R. Koedinger (Co-Chair)
Geoff Kaufman
Roger Azevedo (University of Central Florida)
Emma Brunskill (Stanford University)

Brad A. Myers, Head, Human-Computer Interaction Institute
Martial Hebert, Dean, School of Computer Science

Return to: SCS Technical Report Collection
School of Computer Science homepage

This page maintained by reports@cs.cmu.edu