Abstract
As large language models (LLMs) and vision-language models (VLMs) become increasingly used in robotics area, a crucial question arises: to what extent do these models replicate human-like cognitive processes, particularly within socially interactive contexts? Whilst these models demonstrate impressive multimodal reasoning and perception capabilities, their cognitive plausibility remains underexplored. In this study, we address this gap by using human visual attention as a behavioural proxy for cognition in a naturalistic human-robot interaction (HRI) scenario. Eye-tracking data were previously collected from participants engaging in social human-human interactions, providing frame-level gaze fixations as a human attentional ground truth. We then prompted a state-of-the-art VLM (LLaVA) to generate scene descriptions, which were processed by four LLMs (DeepSeek-R1-Distill-Qwen-7B, Qwen1.5-7B-Chat, LLaMA-3.1-8b-instruct, and Gemma-7b-it) to infer saliency points. Critically, we evaluated each model in both stateless and memory-augmented (short-term memory, STM) modes to assess the influence of temporal context on saliency prediction. Our results presented that whilst stateless LLaVA most closely replicates human gaze patterns, STM confers measurable benefits only for DeepSeek, whose lexical anchoring mirrors human rehearsal mechanisms. Other models exhibited degraded performance with memory due to prompt interference or limited contextual integration. This work introduces a novel, empirically grounded framework for assessing cognitive plausibility in generative models and underscores the role of short-term memory in shaping human-like visual attention in robotic systems.