📄 Full text: arXiv:2605.02703 · local
Most adaptive learning systems are individual-centric and reactive. ProPACT treats collaboration itself as the object of instruction, intervening before dyadic breakdowns occur by forecasting suboptimal states up to 30 seconds in advance.^golrang-propact-pair-programming-2026
The Dyadic Learner Model
ProPACT constructs a real-time model of pair collaboration using three signals:
| Signal | Description | Measurement |
|---|---|---|
| JVA (Joint Visual Attention) | Shared attentional focus | Cosine similarity of gaze distributions over 30-second windows |
| ME (Mental Effort) | Individual cognitive load | Index of Pupillary Activity (IPA) over 10-second windows |
| JME (Joint Mental Effort) | Cognitive engagement synchrony | Cross-recurrence of synchronized ME signals |
All signals are discretized against a resting baseline using ±2SD bins (High, Average, Low).^golrang-propact-pair-programming-2026
Proactive Forecasting Engine
- Model: XGBoost predicts JVA, JME, and ME over a 30-second horizon
- Policy: Rule-based pedagogical hierarchy prioritizes minimally intrusive support
- Principle: Fade support when collaboration is productive; escalate only when breakdown risk increases
Adaptive Feedback Hierarchy
| Intervention | Trigger | Intrusiveness |
|---|---|---|
| A1: Do nothing | MEs=AVG, JVA=H, JME=H | None (desired state) |
| A2: GitHub Copilot | MEs=HH or LL, or (MEs=HL and JVA=L) | Low (autocomplete) |
| A3: Gaze-awareness tool | JVA=Low | Low (visual cue highlighting partner's gaze) |
| A4: Dialogue prompt | JME=Low | Medium (editor prompt encouraging brief dialogue) |
| A5: Task-based hint | Both MEs=High (last resort) | High (directive hint) |
The hierarchy embodies the principle that proactive, minimally intrusive scaffolds outperform reactive, directive interventions.^golrang-propact-pair-programming-2026
Key Results
In a within-subjects study with 26 pair-programming dyads (52 CS/Engineering students), ProPACT feedback vs. control showed:
| Measure | Effect |
|---|---|
| Debugging success | Significantly higher (t(49.96) = −13.51, p < .0001) |
| Debugging time | Significantly lower / more efficient (t(44.70) = 4.39, p < .0001) |
| Feedback uptake | Significantly higher (F(49.81) = −17.69, p < .0001) |
Post-intervention gains in JVA and JME suggest improved collaborative regulation beyond immediate task completion.^golrang-propact-pair-programming-2026
Relationship to Tutoring-Specific Design
ProPACT exemplifies tutoring-specific design: it does not give answers but structures collaboration through graduated scaffolds (mimicking a skilled pair-programming mentor). Like the best tutoring systems, it targets the zone of proximal development at the dyadic level—ensuring partners are neither overloaded nor under-challenged. It also preserves metacognition by requiring learners to actively monitor and regulate their collaboration.^golrang-propact-pair-programming-2026
Implications
- For CS education: Real-time dyadic regulation is tractable and effective; eye tracking + ML forecasting is viable at classroom scale
- For adaptive systems: The "tutoring" unit need not be an individual—collaboration itself can be scaffolded
- For transfer: Does improved dyadic regulation transfer to unsupervised pair work? Not yet tested
Related Pages
- engagement-forecasting-its — Feature-based engagement forecasting reduces MAE 22-33% vs heuristics; effort dr
- clara-collaboration-literacy-dashboard — CLARA: An AI-Augmented Analytics Dashboard for Collaboration Literacy
- ai-agents-peer-learning-discourse — AI agent communities exhibit peer-learning-like discourse patterns
- multi-agent-llm-social-learning — Multi-agent LLMs enable collaborative learning with both AI tutor and AI peers
- genai-meta-analysis-programming-learning — Meta-analytic context for individual ProPACT and tutoring studies
- ai-tutor-behavioral-evaluation — behavioral evaluation axis for AI tutors — measuring what students actually do with feedback
- multimodal-learning-genai — Collaboration as one of six engagement types; group-level cyber-social partnership
- ai-literacy — Collaborative learning with AI partners
- ai-tutor-effectiveness-review — Dyadic regulation as effectiveness factor
- socratic-ai-dialogue — Socratic dialogue in dyadic and group settings
- affective-tutoring — Group-level affect and joint emotional states
- authentic-assessment — Social collaboration and real-time assessment in dyadic contexts
- tutoring-specific-vs-general-ai — Tutoring-specific vs. general-purpose AI design
- ai-k12-evidence-base — Broader evidence landscape
- adaptive-learning-systems — Real-time cognitive-load calibration of dyadic scaffolds
- zone-of-proximal-development — (create when second source emerges)
- programming-its — SCRIPT for Python vs. ProPACT for pair programming
- agentic-education-coding — Agentic vs. collaborative tutoring approaches
- student-perception-ai-use-collaboration -- Misalignment in partners' beliefs about each other's AI use early in a project predicts lower final scores — especially for low-performing teams — and face-to-face sessions don't fix it.
Sources
- Golrang et al. (2026). ProPACT: A Proactive AI-Driven Adaptive Collaborative Tutor for Pair Programming. arXiv:2605.02703. PDF
📎 5 other pages tagged collaborative-ai-tutoring
- AI Literacy
- Beyond the AI Tutor: Social Learning with LLM Agents
- Measuring Cognitive Engagement in Collaborative Discourse with an Extended ICAP Framework: Comparing Human Annotation, In-Context Learning, and Reflective LLM Agents
- Students' Perception Accuracy of Partners' AI Use and its Relation to Collaboration Performance
- When AI Agents Teach Each Other: Discourse Patterns Resembling Peer Learning in the Moltbook Community