<![CDATA[This study addresses the critical need for adaptive learning in non-formal education settings, particularly Community Learning Centres (PKBM) in Indonesia, where student heterogeneity and limited resources challenge conventional teaching methods. We developed a personalized learning path optimization model using Actor-Critic Reinforcement Learning (RL) to enhance STEM competency development. The novel framework integrates cognitive, affective, and personality features to dynamically adjust material difficulty based on real-time analysis of student cognitive states (quiz performance, completion rate) and affective conditions (emotional level), moving beyond static predictive approaches. Experimental results on a synthetic dataset demonstrate that the Actor-Critic agent achieves statistically significant higher rewards (-2.92 vs -3.01, p<0.05) and greater output stability compared to a random baseline. Although the absolute reward difference is modest, it reflects more consistent adaptive policy performance, despite limited effect size (Cohen's d=0.0317). Feature importance analysis confirms that quiz_score and emotion_level are the dominant factors influencing adaptive recommendations, while personality traits show negligible impact. The framework offers a viable pathway for scalable, personalized learning in resource-constrained environments. Future work should validate the model with real-world student data and refine reward functions to strengthen practical impact.]]>
