Approach to Eye Tracking Scanpath Analysis with Multimodal Large Language Model

Eye tracking scanpaths encode the temporal sequence and spatial distribution of eye movements, offering insights into visual attention and aesthetic perception. However, analysing scanpaths still requires substantial manual effort and specialised expertise, which limits scalability and constrains objectivity of eye tracking methods. This paper examines whether and how multimodal large language models (MLLMs) can provide objective, expert-level scanpath interpretations. We used GPT-4o as a case study to develop eye tracking scanpath analysis (ETSA) approach which integrates (1) structural information extraction to parse scanpath events, (2) knowledge base of visual-behaviour expertise, and (3) least-to-most and few-shot chain-of-thought prompt engineering to guide reasoning. We conducted two studies to evaluate the reliability and effectiveness of the approach, as well as an ablation analysis to quantify the contribution of the knowledge base and a cross-model evaluation to assess generalisability across different MLLMs. The results of repeated-measures experiment show high semantic similarity of 0.884, moderate feature-level agreement with expert scanpath interpretations (F1 = 0.476) and no significant differences from expert annotations based on the exact McNemar test (p = 0.545). Together with the ablation and cross-model findings, this study contributes a generalisable and reliable pipeline for MLLM-based scanpath interpretation, supporting efficient analysis of complex eye tracking data.