NAT & NAT-LLM: Adaptive Training + AI Feedback
NAT is a domain-agnostic, multi-parameter adaptive training algorithm that selects each learner's next activity from performance, weaknesses, and cognitive load; NAT-LLM generates after-action feedback from expert-written examples. Applied in a Unity-based naval simulator: NAT statistically outperformed non-adaptive training, and NAT-LLM reached 98% agreement with expert evaluations.
- Unity
- Adaptive Learning
- LLM Feedback
- C#
- Python
The adaptive loop
NAT is a domain-agnostic, multi-parameter adaptive training algorithm: it selects each learner's next activity from performance, weaknesses, and cognitive-load factors rather than a single difficulty knob. NAT-LLM closes the loop by generating after-action feedback from expert-written examples that define the target style and structure of feedback. The model reasons, but the shape of good feedback is fixed by example, not left to the model to invent.
Generations of the work
RAFT (rule-based adaptive difficulty control) → NAT (multi-dimensional weakness adaptation and activity selection) → NAT-LLM (expert-aligned LLM-generated feedback). Each generation kept the adaptive loop above and replaced one piece of it: first the difficulty rule, then the activity-selection policy, then the feedback generator.
Applied and evaluated
Applied in a Unity-based real-time naval simulator for nautical Rules of the Road training, and evaluated through human-subject studies: NAT statistically outperformed non-adaptive training, and NAT-LLM's iteratively refined few-shot feedback reached 98% agreement with expert evaluations across multiple feedback dimensions.
This is the basis of first-author papers at HCII 2025 and HCII 2026 (see Publications) on weakness adaptation and evaluation methodology for adaptive training systems.