2026 ASEE Annual Conference & Exposition

Adaptive tutoring systems increasingly employ reinforcement learning (RL) to personalize instruction, yet many such systems exhibit behaviors that optimize short-term performance at the expense of pedagogically meaningful learning. This work-in-progress paper investigates how instructional behaviors emerge when RL is applied to adaptive tutoring without explicit modeling of knowledge stability and forgetting. Using a simulated tutoring environment framed as a Markov Decision Process (MDP), a tabular State–Action–Reward–State–Action (SARSA) agent selects among three pedagogical actions — teach, quiz, and review — to guide learner progression. The system is evaluated against fixed-sequence and random baselines across 10 independent random seeds using cumulative reward, steps-to-mastery, and instructional action distribution metrics.

Results indicate that while the RL agent converges faster and outperforms both baselines on episodic efficiency — reducing steps-to-mastery from 37.33 (fixed and random baselines) to 25.25 ± 1.13 (aggregated-state) and 26.14 ± 2.02 (full-state) — these efficiency gains reflect optimization of the reward structure rather than pedagogically meaningful learning. The agent converges on a quiz-dominant instructional policy. Across converged policies, quiz actions comprised 36.1–38.6% of selections, with teach and review actions falling below equal distribution. This non-uniform action distribution reveals a structural feature of the instructional environment: in the absence of explicit stability and decay dynamics, quiz actions are reinforced disproportionately by the immediate reward structure (+1 per correct quiz response; +500 terminal mastery bonus; −1 per step), encouraging short-term optimization rather than pedagogically balanced sequencing. Interpreted through Reflexive Reciprocity Theory (RRT), these findings position RRT as a diagnostic framework that uses emergent agent behavior to reveal which pedagogical constructs are operationalized, underspecified, or absent within adaptive instructional designs. The paper contributes design insights for engineering and STEM education by identifying design conditions within this environment under which RL-based tutors can support pedagogically meaningful sequencing, informing the redesign of adaptive systems that balance episodic efficiency with long-term retention.