TutorLoop: Regulating Student Learning Behaviors via Sensor-in-the-Loop Generative Feedback

Synthesis: TutorLoop is a sensor-in-the-loop tutor that reads real-time cognitive states — attention and workload — from ordinary webcam gaze data and uses a deep reinforcement learning agent to choose among four feedback types across a whole learning session, with an LLM refining the wording and tone of the chosen message. Unlike prior AI tutors grounded in course-specific content, TutorLoop takes only sensor-derived signals as input, so a policy trained offline against an LLM-based student simulator could be deployed to new learners and a new task without retraining. In a between-subjects study with N=187 participants, the integrated system produced less frequent but more effective interventions than a control and two baselines, improving attention, lowering workload, sustaining engagement, and raising learning outcomes. The study also found that attention alone weakly predicts learning while workload correlates strongly and negatively, supporting the view that feedback should regulate a cognitive tradeoff rather than maximize a single signal.

Key Findings

1. A DRL agent trained only on simulated students transferred unchanged to a new real task (N=187), suggesting sensor-derived cognitive states generalize better than course-specific content.
2. The Full group combining the DRL model with the LLM tutor reached the highest attention (M=0.739), beating TutorUp (p=.044) and DRL Only (p=.002), while Control scored lower than Full (p=.013).
3. Workload fell furthest in the Full group, significantly below DRL Only (p<.001); the pure DRL policy over-delivered feedback and raised load.
4. TutorLoop delivered significantly less feedback than DRL Only and TutorUp (both p<.001) yet produced stronger outcomes, indicating more selective, adaptively timed interventions.
5. The Full group led on all four outcome metrics (concept M=0.662, detail M=0.448, semantic M=0.284, lexical M=0.151), with significant gains over Control on each (p=.031, p<.001, p<.001, p=.029).
6. Attention correlated only weakly with learning (r=0.017, p=.814) while workload correlated strongly and negatively (r=-0.168, p=.022), echoing the Yerkes–Dodson tradeoff in cognitive psychology.
7. Active engagement via slide switching rose in the Full group (r=0.173, p=.018) even though no feedback explicitly instructed slide switching.

Reading cognitive states from a webcam

TutorLoop captures cognitive states through web cameras, widely accessible and embedded in personal computers, making it suitable for scalable deployment. It uses the Open Source WebGazer library for eye tracking and face detection, running client-side so that raw gaze never leaves the browser — a decision that keeps privacy costs low. Gaze samples are normalized to screen size and split into one-second windows. Attention is the percentage of gaze points inside the primary content area, such as slides or video, following prior student modeling work; workload is gaze entropy, since gaze dispersion tracks cognitive load. Active engagement, measured by slide-switching frequency, is recorded for evaluation only and excluded from model input, because not every task offers comparable button-clicking data. Grounded in the Yerkes–Dodson law, the design treats attention as a proxy for arousal and yields four feedback types: attentive, relief, encouragement, and no feedback.

A DRL agent choosing feedback over a whole session

The core claim is that mapping cognitive states directly to feedback is short-sighted, because intervening whenever attention drops can raise workload and hurt performance. TutorLoop instead trains a deep reinforcement learning agent to optimize feedback across the entire learning session. At each interaction the agent observes a sequence of five units, each a pair of normalized attention and workload values, and picks one of four discrete feedback types. The reward is the relative change in attention minus the relative change in workload against each learner's initial baseline, encouraging the policy to raise attention while holding workload down. Training uses Proximal Policy Optimization with a multilayer perceptron policy, implemented in PyTorch, Stable Baselines3, and Gym, converging at roughly 10,000 steps against the student simulator rather than real learners. This human-in-the-loop setup makes offline training possible for an adaptive learning system.

An LLM tutor that refines the message

The LLM tutor does not choose the action; the DRL agent does. Its role is to refine how the selected action is communicated, countering the habituation effect in which repeated identical messages lose their pull. Working from predefined prompts for each feedback type, the tutor runs a multi-step pipeline: it humanizes the message, contextualizes it using the learner's recent cognitive-state trajectory, and passes it to a tone translator that varies delivery while distinguishing it from previously shown messages. Example tones include concise, friendly, high-energy, mandatory, gentle, playful or humorous, and formal or professional. Variation, perceived monitoring, and humanized phrasing each raise compliance. Crucially, refinement must preserve the intended directive, such as "relax" for relief feedback, and it draws only on sensing data, never course content, so this generative AI layer can transfer across tasks.

The user study and what transferred

The evaluation used a between-subjects design with four groups — Control, TutorUp, DRL Only, and Full — and a final sample of 187 participants (Control 48, TutorUp 46, DRL Only 46, Full 47), from 192 recruited with five excluded for technical failures. Sample size followed an a priori power analysis targeting 80% power to detect a medium effect (0.25) at α=0.05, giving 45 per group. The task was a new self-learning activity: ten Earth-concept slides with three supporting details each, unrelated to the EduAgent materials, studied for 10 minutes and tested for 5 minutes, with feedback appearing every 30 seconds for 10 seconds. WebGazer calibration had to exceed 75% accuracy, and sessions were capped at 15 minutes because tracking drifts beyond that. Crucially, the DRL model was pre-trained on the public EduAgent dataset and applied without retraining, testing transfer of learning across a novel task through quantitative educational measurement of recall.

What this means for practice

• Instructors. Fewer, well-timed prompts beat constant reminders: TutorLoop achieved better outcomes with significantly less feedback, so nudging at moments of declining attention or rising workload matters more than frequency.
• Instructional designers. Treat attention and workload as a tradeoff, not independent targets; interfaces should detect over-engagement and drifting attention, echoing the Yerkes–Dodson account.
• Educational technology developers. A hybrid division of labor works well: let a simulator-trained reinforcement learning policy decide whether and when to intervene, and let an LLM handle phrasing, tone, and personalization.
• Researchers. Domain-agnostic behavioral signals such as gaze can transfer across subjects and platforms without retraining, but cross-context calibration and ethical sensing deserve explicit attention.

Limitations

• The user study, while large at N=187, tested one short self-learning task of ten slides in a 10-minute session, so longer-term learning dynamics remain unobserved.
• Only one task and domain (Earth concepts) was used; the transfer claim rests on offline-to-online application and needs replication across subjects and platforms.
• Cognitive states came from webcam gaze estimation: attention is only a proxy for arousal, WebGazer accuracy required a 75% threshold, and tracking was capped at 15 minutes to avoid drift.
• Offline-to-online transfer depends on the simulator generating a wide enough range of cognitive trajectories, and that simulator is built on the EduAgent dataset rather than real-time student data.
• Outcomes were measured mainly through working-memory recall, with no delayed post-test tracking whether gains persisted.

Citation

Xu, S., & Zhang, X. (2026). TutorLoop: Regulating Student Learning Behaviors via Sensor-in-the-Loop Generative Feedback. arXiv preprint arXiv:2610.09400.