With the increasing accessibility of mobile and digital technologies, video game engagement has significantly increased, leading to emerging concerns around video game addiction. This study addresses the identification of addictive behavioral design elements in video games, which require complex visual reasoning and interpretation ability. We leverage the Qwen-VL-7B Instruct and Llama-3.2-11B-Vision-Instruct vision-language model for this purpose, applying parameter-efficient finetuning (PEFT) using LoRA adapters to tailor it for addiction-related analysis. A custom-curated dataset containing annotated frames from 10 popular video games, each labeled with 8 behaviorally relevant queries, is used for finetuning. The novelty of this research is designing a scheme for the identification of psychologically significant elements in the video, rather than the object detection or action recognition task. Post-training evaluations using metrics such as BERTScore, BLEU, METEOR, and ROUGE show considerable improvements in the model’s ability to align with human-annotated behavioral features. Embedding space analysis further confirms the shift in the model’s internal representation towards human-crafted conceptual understanding. The model shows reliable generalization capability when verified on unseen video game through zero-shot inference. This work paves the way for deploying large vision-language models in the domain of psychological impact assessment through media content. The code and dataset link can be found here: https://github.com/SUNIDHI-SINGH/Video-Game-Addiction-Behavioural-Elements-Dataset .