Difficulties in diagram interpretation in physics: An eye-tracking study of visual attention and diagram interpretation strategies

Analyzing and understanding the differences between novices and experts in domains such as physics has been proven to be important for enhancing competence acquisition in that field. This study investigates these differences through an expert-novice comparison by analyzing visual attention, cognitive strategies, and behavior during scientific diagram interpretation. A total of 23 participants (11 novices and 12 experts) completed 10 diagram-based tasks, while their eye movements and verbalizations were recorded. Experts outperformed novices across several dimensions: higher task success (86% vs 70%), significantly greater visual efficiency (23% vs 17%), and longer dwell times in task-relevant areas. Their shorter, more structured scan paths reflected goal-directed, conceptually driven strategies. In contrast, novices showed disoriented gaze behavior, frequent shifts between the diagram and task text, and relied on surface-level reverse-matching strategies, indicating higher extraneous cognitive load. Think-aloud data revealed qualitative reasoning differences: experts used precise, technical language and referenced diagrams accurately, while novices were vague, hesitant, and uncertain. Emotional expressions in novice protocols aligned with signs of cognitive overload. These findings support cognitive load theory, suggesting experts draw on schemas to manage intrinsic load, while novices struggle with element interactivity. Unexpectedly, some experts in the study showed novicelike behavior, whereas some novices exhibited patterns typically associated with expert performance. Novicelike experts often lacked recent exposure to diagram-related coursework. Conversely, some novices demonstrated efficient gaze patterns and structured scanning but still made conceptual errors. To contextualize human performance, two large language models (ChatGPT-4 and DeepSeek-R1) were tested on the same tasks. ChatGPT-4 reached an overall success rate of 50%, handling simple extractions but failing with visual relationships. DeepSeek-R1 performed poorly (10%), revealing current AI limitations in visual interpretation. These results underscore the uniquely human capacity for visual reasoning, schema activation, and metacognitive monitoring, which AI still lacks. The findings emphasize the critical role of graph literacy, representational competence, and cognitive strategy use in diagram interpretation. They highlight the need for explicit instructional interventions to help novices develop expertlike visual attention patterns and reasoning skills.