Towards an adaptive AI scaffold for developing student collaborative problem solving

Synthesis: Wong, Bulathwela, and Cukurova study how to design an adaptive AI scaffold for developing students' collaborative problem solving (CPS) that responds to the dynamic nature of individual students' processes. Unlike most adaptive-scaffolding designs — which are informed by student choice or rely on aggregated behavioural/performance indicators — they derive scaffold strategies from sequential patterning mining of individual students' process sequences. Using speech and task logs from 78 students (aged 14–15) working in triads on a mathematics CPS task via online video conferencing, they found that while students with a maximal scaffold achieved greater performance improvements and were significantly more on-task than those with a minimal scaffold, they also engaged in more scripting behaviours. The derived adaptive scaffold design informs conceptual strategies for the 'problem identification' and 'ideation, planning and decision making' phases of CPS.

Key Findings

• Adaptive scaffolding is needed but rarely process-dynamic. Most adaptive-scaffold designs are informed by student choice or aggregated behavioural/performance measures, rather than the dynamic, individual-level process sequences of students during CPS.
• Sample & task. 78 students aged 14–15 from a public school worked in triads on a mathematics collaborative problem-solving task through online video conferencing.
• Method: sequential patterning mining. The adaptive scaffold design was derived using sequential patterning mining with independent-effects logistic regression to identify sequence-based patterns associated with students' performance improvements on a mathematics problem similar to the CPS task.
• Maximal vs minimal scaffold. Students with the maximal scaffold had greater performance improvements and were significantly more on-task than those with the minimal scaffold — but they also engaged in more scripting behaviours.
• Phase-specific strategies. The derived design informs conceptual strategies such as prompting students who are capable of answering questions to ask questions in the 'problem identification' phase.
• Key CPS phases. The design highlights specific strategies for the 'problem identification' and 'ideation, planning and decision making' phases.

Implications

This work connects adaptive learning, scaffolding, and collaborative learning for secondary-school students, using process-mining learning analytics to move adaptive scaffolding from static/aggregate rules toward individualized, sequence-based intervention. The finding that maximal scaffolding boosts on-task performance but increases scripting behaviours is a nuance for engagement and agency debates: scaffolding that keeps students on-task may also script their interactions. The proposed adaptive scaffold — prompting capable students to ask questions during problem identification — offers a concrete, process-grounded design for AI-assisted collaborative learning in mathematics.

Connected Concepts

• Collaborative Learning — the CPS context under study
• Scaffolding — the adaptive scaffold being designed
• Adaptive Learning — the personalization approach
• Learning Analytics — process-mining methodology on speech/task logs
• Student Modeling — individual-level process modeling
• K 12 — the secondary-school sample
• Student Engagement — on-task vs scripting behaviours
• Metacognition — the phases (problem identification, ideation/planning) scaffolded
• Self Regulated Learning — individual regulation within collaborative work

Citation

Wong, K., Bulathwela, S., & Cukurova, M. (2026). Towards an adaptive AI scaffold for developing student collaborative problem solving. Learning and Instruction, 105, 102418.