Towards Real-Time Personalized Feedback in Open-Ended Learning Environments

Open-ended learning environments (OELEs), such as online science labs, enable students to explore and develop inquiry strategies through educational exercises not available in real-life. Unfortunately, students often struggle to effectively learn in these environments. Prior research has mainly focused on student profiling and already established inquiry strategies. In this work, we introduce a novel approach in which a reinforcement learning agent is trained with additional behavioral cloning loss to act as a realistic, high-performing student within an OELE. At deployment, the agent identifies optimal actions, which are then transformed via LLM into personalized feedback messages, which are then presented to students. We evaluate our method using data from an existing OELE and in a real-life classroom study on 103 students. Our results show that our agent was able to identify action sequences that lead to high performance, and that students who followed their personalized feedback developed useful, general inquiry strategies. This work establishes a foundation for implementing real-time personalized feedback in any open-ended, exploratory setting with clickstream data and a clear performance measure, without relying on prior notions of effective inquiry.