Learning to Prompt: Improving Student Engagement with Adaptive LLM-based High-School Tutoring

Researchers developed an adaptive large language model tutoring system that uses subject-aware prompting and machine learning to personalize education for high-school students. Testing with 656 conversations showed the system improved instructional efficiency by reducing interactions by ~3 turns and increased exercise completion rates to 28.1% using stochastic strategy sampling, demonstrating effective sim-to-real transfer from simulation training to live student interactions.

This research addresses a fundamental challenge in AI-assisted education: static prompting strategies fail to accommodate diverse academic disciplines and individual learning needs. The team engineered a prompt routing model trained in simulation before deployment with real students, enabling dynamic selection between analytical and scaffolding-based learning strategies depending on student performance and comprehension levels.

The work builds on growing recognition that LLM effectiveness depends heavily on prompt engineering and contextual adaptation. Traditional tutoring systems apply fixed instructional approaches regardless of subject matter or student progress, limiting their pedagogical value. By extracting 14 pedagogical features from transcripts—including tutor scaffolding quality and student understanding—the researchers created measurable signals for intelligent strategy switching.

The empirical results demonstrate meaningful real-world impact. The stochastic router achieved a 28.1% exercise conversion rate versus 19.6% for static baselines, suggesting students benefit from varied instructional approaches rather than monotonous consistency. Reducing interaction length by 3 turns while maintaining educational quality addresses practical classroom constraints where instruction time is finite.

This advance matters beyond academia. As organizations explore AI-driven customer support and training systems, adaptive prompting patterns directly transfer. The successful sim-to-real transition validates that training environments can effectively prepare models for production deployment. Future developments likely involve incorporating real-time performance metrics and expanding to specialized domains where precise pedagogical tailoring becomes increasingly valuable.

• →Adaptive LLM prompting outperforms static baselines by 7.3% on simulation benchmarks and reduces instructional interactions by 3 turns in live testing
• →Stochastic strategy sampling achieved 28.1% exercise completion rates versus 19.6% for static approaches, indicating diversity in instruction improves outcomes
• →14 pedagogical features extracted from transcripts enable intelligent switching between analytical and scaffolding learning strategies based on student needs
• →Successful sim-to-real transfer validates that simulation-trained prompt routing models generalize effectively to actual student interactions
• →Findings have implications for AI-driven customer support and training systems requiring dynamic contextual adaptation