ProPACT: Pair Programming with AI

ProPACT (Proactive AI-Driven Adaptive Collaborative Tutor) is an AI-driven adaptive tutoring system for pair programming that treats collaboration itself as the object of instruction. Unlike individual-centric, reactive systems, it models dyadic learning states in real time and intervenes before collaborative breakdowns occur, using multimodal sensing and predictive forecasting.

Key Findings

1. Significant performance gains from proactive feedback. Dyads receiving ProPACT feedback achieved substantially higher debugging success (t[49.96] = −13.51, p < .0001) and completed tasks more efficiently (t[44.70] = 4.39, p < .0001) compared to the no-feedback control condition.

2. Dyadic sensing enables predictive intervention. ProPACT constructs a multimodal dyadic learner model from Joint Visual Attention (JVA — cosine similarity of gaze distributions over 30-second windows), Joint Mental Effort (JME — cross-recurrence quantification of pupil-diameter signals), and individual Mental Effort (IPA from pupillary fluctuations). An XGBoost-based forecaster predicts sub-optimal collaboration states up to 30 seconds in advance.

3. Five-tier adaptive feedback hierarchy works. The system escalates through minimally intrusive scaffolds: (A1) do nothing when collaboration is productive; (A2) temporarily enable GitHub Copilot when cognitive strain rises; (A3) show a gaze-awareness tool highlighting the partner's visual focus; (A4) issue unobtrusive dialogue prompts to re-align mental effort; and (A5) provide directive task-based hints only as a last resort. Signals are discretized against a normalized resting baseline using a ±2SD criterion (High / Average / Low).

4. Post-intervention gains in collaborative regulation. Beyond task-level improvements, dyads showed sustained increases in JVA and JME after the intervention, indicating that the system fostered durable collaborative skills rather than just providing momentary assistance.

Implications

ProPACT represents a shift from individual to dyadic learner modeling in Intelligent Tutoring. By treating the pair — not the person — as the unit of analysis, it addresses a long-standing gap in Collaborative Learning support. Traditional ITS architectures focus on individual cognition; ProPACT demonstrates that multimodal signals (gaze, pupil dilation) can be fused to model the health of a collaborative process in real time.

The proactive forecasting approach is a departure from reactive feedback paradigms common in Adaptive Learning. By predicting breakdowns 30 seconds ahead, ProPACT avoids the latency inherent in "detect-then-respond" architectures, allowing scaffolds to arrive before students experience frustration or disengagement. This has implications for engagement-metrics and real-time classroom orchestration.

For CS Education specifically, ProPACT validates that AI-assisted pair programming can improve both task outcomes and collaborative skill development. The system's integration with Collaborative AI Tutoring workflows suggests a future where AI tutors monitor not just what students produce (code), but how they work together.

The gaze-awareness tool (A3) is a particularly novel intervention: rather than providing didactic content, it surfaces the partner's attentional focus as a lightweight nudge toward shared attention. This aligns with Multimodal AI Tutoring research emphasizing non-verbal channels for learning support.

Connected Concepts

Adaptive Learning

AI Tutoring

RAG

Connected Articles

Collaborative AI Tutoring — Collaborative AI Tutoring

Multimodal AI Tutoring — Multimodal AI Tutoring in STEM

Adapt Adaptive Lesson Plan Transformer — AdaPT: Adaptive Lesson Plan Transformer for Cross-Regional and Differentiated Instruction

Adaptive Pretesting Retention — Do Gains from Generative AI-Enabled Adaptive Pretesting Persist? Evidence from a Retention Study

Adhd Video Segmentation Computing Education — Leveling the Playing Field: Temporal Video Segmentation for Individuals with ADHD in Computing Education

Agents That Teach Incidental Learning — Agents That Teach: Designing Incidental Learning Back into AI-Assisted Software Development

AI Adult Learning Design — Guidelines for Designing AI Technologies to Support Adult Learning

AI Agents Constructive Conflict Design Education 2026 — Enacting Constructive Conflicts with AI Agents to Enhance Reconsideration among Novice Interaction Designers

AI Coaching RL Skill Development — AI Coaching for Accelerating Human Skill Development with Reinforcement Learning

AI Enabled Serious Games — AI-Enabled Serious Games: Integrating Intelligence and Adaptivity in Training Systems

AI Generated Traces Novice Programmers — AI-Generated Traces for Novice Programmers: Learning Effects and Learner Differences in a Multi-Institutional Study

Stanford Evidence Base AI K12 2026 — AI in K-12 Evidence Base

AI Learning Companions Framework — Building AI Companions that Prioritise Learning over Performance

Transfer Of Learning — AI Learning Transfer

AI Powered Personalized Learning Elementary Fractions 2026 — Exploring Fraction Comprehension and Interest in Elementary Education Through AI-Powered Personalized Learning

AI STEM Bibliometric Trends — Why does AI unlock new possibilities in STEM education? A Bibliometric Analysis of Trends and Future Agenda

AI Tutor Safety Harms — AI Tutor Safety and Pedagogical Harms

AI Vocational Education Training Review — Artificial intelligence in vocational education and training: A systematic review of educational purposes, theoretica...

Anvil AI Educational Animations — ANVIL: Analogies and Videos for Lecturers

Astra Atco Training Simulator — ASTRA: A Scalable Next-Generation ATCO Training Simulator with Autonomous Simpilots

Automated Grading Linux Bash Examinations Large Language Models — Automated Grading of Linux/Bash Examinations Using Large Language Models

Behaviorally Adaptive Visual Diversion Assessment 2026 — Behaviorally Adaptive Visual Diversion for Inclusive and Resilient Digital Assessment Delivery

Buggy GenAI Code Student Responses — When AI Is Wrong on Purpose: How Students Respond to Buggy GenAI Code

Chat Debugging LLM Hardware Education 2026 — WIP: Chat-Debugging: Large Language Model as a Hardware Debugging Assistant

Clara Collaboration Literacy Dashboard — CLARA: An AI-Augmented Analytics Dashboard for Collaboration Literacy

Citation

Golrang, A., Sharma, K., Dehaen, S., & Viberg, O. (2026). ProPACT: A Proactive AI-Driven Adaptive Collaborative Tutor for Pair Programming. arXiv:2605.02703.