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

Abstract: LLMs can personalize education, although current static-prompt tutoring systems struggle to adapt to diverse academic disciplines. We develop and test a system with subject-aware prompting, based on 14 pedagogical features (e.g., tutor scaffolding, student understanding) extracted from raw transcripts. We first train a prompt routing model in a simulation environment, and then deploy it for online adaptation with actual high-school students. The simulation benchmark shows the router outperforming two static baselines ($0.694$ vs. $0.647$ and $0.64$, $p<0.001$). A/B testing ($N=656$ conversations from 359 students) shows sim-to-real transfer where the model switches from analytical to scaffolding learning strategies. Our adaptive prompt selection mechanism improves instructional efficiency, maintains pedagogical quality and reduces interactions by around 3 turns ($p=0.007$). While a greedy router achieves a comparable exercise conversion rate with the baseline ($19.1\%$ vs. $19.6\%$), a stochastic router that samples strategies leads to a higher conversion rate ($28.1\%$).

• The paper presents a novel subject-aware prompt routing system that dynamically selects expert-authored pedagogical strategies based on contextual student-tutor interactions.
• It employs a dual-path embedding architecture and a contextual bandit formulation to optimize safety-controlled prompts and mitigate representation anisotropy.
• Empirical evaluations in both simulated and real-world settings demonstrate improved instructional efficiency, reduced session redundancy, and higher student conversion rates.

Subject-Aware Adaptive Prompt Routing for LLM-Based High-School Tutoring

Introduction

The paper "Learning to Prompt: Improving Student Engagement with Adaptive LLM-based High-School Tutoring" (2606.20138) presents a paradigm-shift in leveraging LLMs for education by introducing a subject-aware, adaptive prompt routing framework. The authors critically analyze limitations of static prompt-based tutoring systems, notably their lack of cross-disciplinary adaptability and inability to bridge simulation-reality performance gaps, and propose a solution that dynamically selects pedagogical strategies based on rich context features extracted from student-tutor dialogues.

Methodological Contributions

Contextual Bandit Formulation and Architecture

The core of the framework is the reformulation of adaptive pedagogical strategy selection as a contextual bandit prompt routing problem. The routing model observes a hybrid subject-topic embedding, constructed via a fused architecture combining a pre-trained E5 encoder with a learnable subject ID embedding. This dual-path mechanism effectively counteracts the representation anisotropy and semantic collapse typical of large-scale pre-trained embeddings, especially when crossing disciplinary boundaries.

The action space comprises 20 safety-constrained, expert-authored pedagogical prompts, each aligned with a distinct instructional strategy, such as Socratic questioning, scaffolding, Feynman inversion, and affective support. These prompts form a discrete set of policies, enforce pedagogical best-practices, and enable direct interpretability and safety control.

The router's architecture leverages an actor-critic network with a residual subject embedding injection, ensuring strong subject-conditionality in the final policy head and mitigating cross-task interference.

Feature-based Pedagogical Evaluation

A critical innovation is the construction of a high-dimensional, feature-based pedagogical feedback signal. The authors decompose instructional efficacy into 14 criteria—ranging from scaffolding and diagnostic behavior to learner curiosity and mistake justification—using a GPT-5-mini LLM evaluator calibrated on human-annotated ground truth. Weighted aggregation of these binary features produces a robust pedagogical score, empirically aligned with downstream exercise conversion.

To bridge the simulation-reality distributional shift, the authors calibrate the feedback signal via sigmoid temperature scaling (K=3), minimizing Jensen-Shannon divergence between simulated and empirical score distributions.

Optimization and Deployment

Policy optimization is performed with PPO, truncated to single-step contextual bandit mode, augmented by experience replay and importance sampling for sample efficiency—critical in the high-latency, sparse-reward tutoring environment.

In live deployment, a policy mixture strategy is used: 80% greedy exploitation to ensure robust user experience, 20% exploration via stochastic prompt sampling to preserve policy diversity and enable ongoing on-policy data collection. Continuous online finetuning of the router is implemented via real-world training data.

Empirical Results

Simulation and Robustness

The dynamic router yields statistically significant improvements over both a commercial production baseline (mean pedagogical score 0.640) and a state-of-the-art static literature prompt (0.647), achieving a mean score of 0.694 ($p < 0.001$). Particularly noteworthy is the elimination of the lower tail of underperforming interactions; the system demonstrates resilience even with simulated “unmotivated” students, outperforming static prompts especially for low-engagement profiles.

Empirical evaluation of subject-strategy mapping shows that in 8 out of 13 subjects, the routing model converges to the optimal empirical prompt (within a regret gap $<0.01$), with model selection accuracy of 69.2% in simulation. The regret margins on sub-optimal subject-prompt pairings are minimal.

Real-World Deployment and Sim-to-Real Transfer

A/B testing on 656 conversations with 359 high-school students establishes a sim-to-real transfer effect. The router maintains pedagogical parity with the production baseline (mean scores 0.565 vs. 0.569, $p = 0.635$), but crucially, over the deployment window, it refines its strategy selection in response to real user feedback. A policy shift is observed: simulation-favored strategies (e.g., Feynman explanation) are superseded by routings toward scaffolded support (Coach) in response to empirical dialog dynamics.

Notably, the router increases instructional efficiency: in sessions of depth (≥5 message turns), it reduces session length by approximately 3 full turns compared to baseline ($p = 0.007$), indicating lower instructional redundancy while maintaining learning progression.

Downstream Learning Engagement

Higher LLM-evaluated pedagogical scores correlate with increased student transition to formal exercises (conversion rate: 0.599 vs. 0.560, $p = 0.037$). While the greedy router achieves comparable conversion with baseline (19.1% vs. 19.6%), exploration via stochastic prompt selection elevates conversion to 28.1%—a strong, albeit not statistically significant, signal that policy diversity enhances engagement.

Within-user analysis shows a non-significant trend ($p=0.281$) towards higher exercise accuracy after high-quality tutoring interactions, suggesting that while engagement is immediately measurable, direct knowledge accrual may be longitudinal.

Theoretical and Practical Implications

This work substantiates several claims:

• Subject-aware prompt routing outperforms static prompting in both pedagogical feedback and engagement, particularly by mitigating the variance and failure rate associated with static, one-size-fits-all policies.
• Decomposition of instructional effectiveness into fine-grained, observable behavioral criteria enables stable, robust reward signals, overcoming the pitfalls of black-box reward gaming and variance in scalar feedback.
• Sim-to-real adaptation is feasible and essential; subject-adaptive routers appropriately abandon simulation priors in favor of empirically validated, student-aligned strategies post-deployment.
• Exploration-induced diversity in instructional strategy increases engagement—favoring policy mixtures over pure exploitation, a result with implications for continuous RL in sparse-feedback, high-variance user-facing environments.

Practically, the framework offers an interpretable, safe, and extensible architecture for deploying LLM-based tutors across diversified curricula, with evidence of tangible gains in both efficiency and student engagement metrics.

Limitations and Future Directions

Several limitations are acknowledged. The current discrete prompt pool, while ensuring pedagogical safety, constrains the expressivity of possible strategies, limiting spontaneous hybrid or novel instructional tactics. Data sparsity and domain imbalance in real-world deployment restrict rapid convergence on subject-specialized optima, especially in tail subjects with few interactions.

Immediate assessment of learning gain is hampered by low exercise conversion rates after tutoring; thus, measuring long-term knowledge retention remains an open challenge. The model does not utilize full conversational history, potentially curtailing finer session-level adaptation.

Potential research avenues include integration of conversation-level contextual memory, e-greedy/bayesian optimization for more principled exploration-exploitation balance, and dynamic prompt generation/expansion beyond finite pools while maintaining safety guarantees.

Conclusion

The proposed subject-aware, adaptive prompt routing system for LLM-based high-school tutoring (2606.20138) empirically demonstrates improved instructional efficiency and engagement compared to static approaches, both in controlled simulation and real-world classroom deployment. By fusing robust feature-based evaluative feedback with architecture-level innovations in subject-context representation, the framework provides a powerful mechanism for contextually optimal, pedagogically valid, and safe LLM-driven education systems. Continuing advances in adaptive policy modeling, feedback calibration, and session-aware strategy selection will be crucial for realizing the full potential of LLMs in open-domain, personalized, and scalable education.