Tested on 187 participants, the system intervenes less frequently but more appropriately, resulting in improved concentration, reduced mental load and better results.
Tutors based on large language models react mainly to the content of the lesson. TutorLoop, presented by Songlin Xu and Xinyu Zhang, starts with another signal: the learner's cognitive state, estimated in real time from the webcam.
The system works in a closed loop. A deep reinforcement learning agent chooses the type of feedback to give, optimizing the entire session rather than reacting mechanically to each drop in concentration. A second language model then transforms this choice into a natural message, adapted to the context.
The evaluation involved 187 people, with a model trained offline and then applied without retraining to a new learning task. The result: rarer but more effective interventions that improve attention, reduce perceived workload, increase engagement and, ultimately, achievements.
Read the full article on the original source
Go to the source →