A Survey on Emotion-Aware Programming Assistance Using Large Language Models, Eye-Gaze Analysis and Keystroke Dynamics

Large language models (LLMs) are entering programming education quickly, where they are used to produce hints, feedback, exercises, and code explanations. Most of these tools respond only to the code or error message in front of them and remain unaware of whether the learner is confused, frustrated, bored, or overloaded. This survey reviews twenty studies that bear on this problem, covering LLM-based programming assistance, eye tracking and gaze-aware assistants, facial-expression recognition of engagement and frustration, keystroke dynamics, and behavioral indicators of cognitive load. The literature shows that LLM feedback is generally well received and becomes more useful when it is tied to the student’s own code, that gaze reflects attention and comprehension effort, and that typing rhythm changes with emotional state and can be captured without a camera. However, most studies examine a single signal, rely on controlled or non-programming tasks, depend on self-reports, or target authentication rather than learning. From this synthesis we identify a gap: we found very little work that links learner-state sensing to LLM feedback in programming, and none that tests whether doing so improves learning. We outline a conceptual emotion-aware programming tutor as a research direction, kept separate from the reviewed evidence, and discuss privacy, bias, and evaluation challenges.