[Paper] Orchestrating Attention: Bringing Harmony to the 'Chaos' of Neurodivergent Learning States

Source: arXiv - 2602.07865v1

Overview

The paper introduces AttentionGuard, a novel adaptive‑learning framework that senses a learner’s moment‑to‑moment attention state and reshapes the user interface to keep neurodivergent users—especially those with ADHD—engaged without overwhelming them. By turning “attention chaos” into a predictable signal, the authors show how subtle UI tweaks can lower cognitive load and boost comprehension in real‑time learning environments.

Key Contributions

• Four‑state attention model grounded in ADHD research (under‑stimulated, focused, overstimulated, disengaged).
• Five UI adaptation patterns (e.g., bi‑directional scaffolding, dynamic pacing, visual‑noise modulation) that react to both low and high arousal.
• Privacy‑preserving detection pipeline using only coarse behavioral cues (mouse movement, keystroke timing, viewport focus).
• Robust classifier achieving 87.3 % accuracy on the large‑scale OULAD dataset and validated against clinical ADHD profiles (HYPERAKTIV).
• Wizard‑of‑Oz user study (N = 11) demonstrating significant reductions in perceived workload (NASA‑TLX: 47.2 vs 62.8) and higher comprehension scores (78.4 % vs 61.2 %).
• Open demo that visualizes detected states, UI adaptations, and human‑in‑the‑loop overrides, enabling reproducibility and further experimentation.

Methodology

1. Signal Collection – The system logs lightweight, non‑identifying interaction data (mouse velocity, click intervals, scrolling speed, window focus changes).
2. Feature Engineering – Temporal windows (2‑second slices) are transformed into statistical descriptors (mean, variance, entropy) that capture rapid fluctuations typical of ADHD attention patterns.
3. State Classification – A shallow ensemble (Random Forest + Gradient Boosting) is trained on the OULAD dataset, where ground‑truth labels were derived from self‑reported focus surveys. Cross‑validation on the independent HYPERAKTIV clinical dataset confirms generalizability.
4. UI Adaptation Engine – The detected state triggers one of five pre‑designed UI “recipes”:
   • Stimulus Boost: add interactive prompts when under‑stimulated.
   • Noise Dampening: simplify visual clutter when overstimulated.
   • Pacing Slider: adjust content delivery speed based on engagement level.
   • Bi‑directional Scaffolding: provide hints for disengagement and allow self‑pacing for hyper‑focus.
   • Feedback Loop: surface a subtle visual cue indicating the system’s current inference, letting users override if needed.
5. Evaluation – A Wizard‑of‑Oz study compared the adaptive UI against a static baseline. Cognitive load (NASA‑TLX), comprehension quizzes, and agreement between human wizards and the classifier were recorded and statistically analyzed.

Results & Findings

The findings suggest that detecting attention states from innocuous interaction data is feasible and that responsive UI changes can materially improve learning outcomes for neurodivergent adults.

Practical Implications

• EdTech Platforms can embed AttentionGuard‑style modules to automatically adjust lesson pacing, visual density, or interactive prompts, reducing the need for manual accommodations.
• Enterprise Training tools gain a “smart tutor” layer that keeps employees with attention variability on track, potentially lowering dropout rates in compliance courses.
• Assistive Technology developers obtain a privacy‑first blueprint for building attention‑aware interfaces without requiring eye‑tracking hardware or invasive biometrics.
• Product Designers can adopt the five UI patterns as design primitives, applying them to dashboards, code editors, or collaborative whiteboards where focus swings are common.
• Data‑Governance Teams benefit from the demonstrated feasibility of using only aggregate behavioral signals, easing GDPR/CCPA compliance concerns.

Limitations & Future Work

• Sample Size & Demographics – The user study involved a small, adult‑only cohort; results may differ for children or for other neurodivergent conditions (e.g., autism).
• Signal Granularity – Relying on coarse mouse/keyboard data may miss subtler physiological cues; integrating optional eye‑tracking or heart‑rate data could improve accuracy.
• Generalization Across Domains – The classifier was trained on a MOOC dataset; transferring to highly interactive environments (e.g., VR, gaming) will require domain‑specific retraining.
• Long‑Term Adaptation – The current system reacts per 2‑second window; future work should explore cumulative learning models that adapt to evolving user strategies over weeks or months.

The authors propose expanding the attention taxonomy, testing in K‑12 settings, and open‑sourcing the detection pipeline to accelerate community‑driven refinements.

Authors

• Satyam Kumar Navneet
• Joydeep Chandra
• Yong Zhang

Paper Information

• arXiv ID: 2602.07865v1
• Categories: cs.HC, cs.AI
• Published: February 8, 2026
• PDF: Download PDF