Conversational Agents for Adaptive Learning: Generative Instruction, Reinforcement Fine-Tuning, and Behavioral Optimization in Digital Education

This study introduces a conversational adaptive learning agent that integrates generative language modeling with reinforcement fine-tuning to optimize instructional behavior in real-time learner interactions. The system was trained on 39,000 interaction turns from 2,300 training sessions involving 520 learners, validated on 9,200 turns from 540 sessions, and tested on 10,800 turns from 610 sessions collected across mathematics, science, language, and computer-science domains. Offline evaluation compared a baseline generative model, a correctness-only reinforcement model, and a composite-reward configuration integrating correctness, engagement, and sentiment. The composite-reward model achieved a pedagogical adequacy rating of 4.3 versus 3.4 for the baseline, reduced safety violations from 14.2 to 8.9 per 1,000 turns, and lowered average turn length from 62.5 to 55.1 tokens while improving clarity scores from 3.8 to 4.1 (1–5 scale). A three-week online user study involving 84 participants demonstrated significant behavioral differences. Average session duration for the reinforcement-tuned group increased from 25.4 to 28.6 minutes, while the baseline group declined from 21.3 to 19.8 minutes. Voluntary session extension rose from 14% to 33%, hint-seeking frequency increased from 0.22 to 0.36, and frustration indicators dropped from 19% to 11%. Progression into higher-difficulty tasks reached 44% by Session 5, compared to 21% for the baseline. The study positions reinforcement-tuned conversational agents as a viable direction for scalable, affect-sensitive adaptive learning.