ProPACT: A Proactive AI-Driven Adaptive Collaborative Tutor for Pair Programming

Synthesis: ProPACT (Proactive AI-Driven Adaptive Collaborative Tutor) is an adaptive tutoring system for pair programming that treats collaboration itself as the object of instruction. Unlike individual-centric, reactive systems, it builds a real-time model of dyadic learning from Multimodal AI sensing and intervenes before collaborative breakdowns occur. Three signals define that model: Joint Visual Attention (JVA), Joint Mental Effort (JME), and individual Mental Effort (ME), each discretized against a resting baseline using a ±2SD criterion. An XGBoost forecaster predicts sub-optimal collaboration states up to 30 seconds ahead, driving a five-tier Scaffolding hierarchy that escalates from doing nothing to a directive hint only as a last resort. In a within-subjects study with 26 dyads (52 CS and engineering students), ProPACT feedback produced higher debugging success, faster debugging, and greater feedback uptake than a no-feedback control, with post-intervention gains in JVA and JME suggesting durable collaborative AI Regulation in Education. The work reframes the unit of tutoring as the pair rather than the person.

Key Findings

1. Proactive feedback lifts performance. Dyads receiving ProPACT feedback achieved higher debugging success (t[49.96] = −13.51, p < .0001) and finished debugging faster (t[44.70] = 4.39, p < .0001) than the no-feedback control.
2. Dyadic sensing enables prediction. JVA (gaze-distribution cosine similarity over 30-second windows), JME (cross-recurrence of pupil signals), and ME (Index of Pupillary Activity) feed one model that forecasts sub-optimal states 30 seconds ahead.
3. A five-tier hierarchy delivers minimally intrusive help. Escalation runs A1 (do nothing), A2 (GitHub Copilot), A3 (gaze-awareness cue), A4 (dialogue prompt), A5 (directive hint), used only when signals show rising breakdown risk.
4. Collaborative skill gains persist. Beyond immediate task outcomes, dyads showed post-intervention increases in JVA and JME, indicating that the system scaffolded durable collaborative AI Regulation in Education rather than providing momentary assistance.
5. Support is graduated, not answer-giving. ProPACT mimics a skilled pair-programming mentor, structuring collaboration within each pair's zone of proximal development rather than giving solutions, and partners must monitor their own Metacognition.
6. Boundaries of the evidence. The 52-student within-subjects design cannot show whether better dyadic regulation transfers to unsupervised pair work, and deployment depends on eye-tracking hardware, limiting Learning Gains claims.

The Dyadic Learner Model

ProPACT treats the pair, not the individual, as the unit of learner modeling. Three signals are captured continuously during pair programming and discretized against a normalized resting baseline into High, Average, and Low bins using a ±2SD criterion:

Fusing gaze and pupillary channels lets the system estimate how well a pair coordinates attention and effort — a signal unavailable to single-learner adaptive systems.

Proactive Forecasting and the Feedback Hierarchy

An XGBoost model predicts JVA, JME, and ME over a 30-second horizon, and a rule-based pedagogical hierarchy converts those forecasts into the least intrusive support warranted, fading help when collaboration is productive and escalating only when risk rises:

The gaze-awareness tool surfaces the partner's attentional focus rather than supplying content, aligning with work on non-verbal support channels. The tiers embody a Feedback principle: proactive, minimally intrusive nudges beat reactive correction.

Key Results

In a within-subjects study with 26 pair-programming dyads (52 CS and engineering students), ProPACT feedback was compared against a no-feedback control. Feedback-condition dyads solved debugging tasks more often and more quickly, and followed the system's suggestions more readily (uptake: F[49.81] = −17.69, p < .0001). Post-intervention gains in JVA and JME point to improved collaborative AI Regulation in Education beyond immediate task completion, suggesting the scaffolding left a residue of better joint attention rather than merely speeding up one task. The authors read this as evidence that real-time dyadic regulation is tractable in realistic programming settings and that engagement and effort synchrony can be monitored without interrupting the work itself. Gains held for both outcome and process measures.

Tutoring-Specific Design and Implications

ProPACT exemplifies tutoring-specific design in the sense of evidence-based tutoring: it withholds answers and instead structures collaboration through graduated Scaffolding, positioning the system as a pedagogical agent rather than a solution dispenser. Metacognition survives because partners must notice and repair their own coordination, and the sociocultural framing places the zone of proximal development at the level of the dyad.

For CS Education, the study shows that real-time dyadic regulation is tractable and effective, and that eye tracking plus machine-learning forecasting is viable at classroom scale. For adaptive systems more broadly, the "tutoring" unit need not be an individual — collaboration itself can be scaffolded, complementing reviews such as A systematic review of AI-powered collaborative learning in higher education: Trends and outcomes from the last decade. What remains untested is transfer: whether improved dyadic regulation carries into unsupervised pair work, where no tutor is watching.

What this means for practice

• Instructors. Scaffold the pair rather than the person: monitor joint visual attention and joint mental effort and intervene before coordination breaks down, which raised debugging success (t[49.96] = −13.51) and speed.
• Instructors. Escalate help in tiers and begin from doing nothing — gaze cues and dialogue prompts before any directive hint — so partners retain the coordination work themselves.
• Designers. Treat collaborative state as a forecastable signal: an XGBoost model over JVA, JME, and individual mental effort predicted sub-optimal states 30 seconds ahead.
• Designers. Test whether cheaper signals such as interaction logs, keystroke dynamics, or webcam gaze can carry the forecasting model before planning classroom deployment, since dual eye tracking is the scaling constraint.
• Instructors. Verify that improved joint attention persists into pair work you do not supervise; this study could not test that transfer.

Limitations

• Twenty-six dyads (19 female, 33 male) of undergraduate and master's students in computer science or engineering at a single European university.
• The study ran in a controlled laboratory setting with short-duration debugging tasks, which the authors say may not generalize to classroom, remote, or industrial contexts.
• The system depends on specialized sensing infrastructure — dual eye tracking and pupillometry — which poses practical and financial constraints for large-scale deployment.
• Tasks contained only logical bugs and no syntax errors, so the forecaster has not been tested on open-ended or ambiguous programming work, and transfer to unsupervised pair work is untested.

Citation

Golrang, A., Sharma, K., Dehaen, S., & Viberg, O. (2026). ProPACT: A Proactive AI-Driven Adaptive Collaborative Tutor for Pair Programming.