Abstract: To address the limitation of traditional packaging design methods in capturing users’ personalized aesthetic preferences, this study proposes a cigarette packaging design preference analysis method that integrates a Visual-Language Model (VLM) with a Large Language Model (LLM), and constructs a multi-agent collaborative framework named Box Insight. First, online reviews of cigarette packaging were collected through web crawling, and Kansei-related vocabulary reflecting consumer aesthetic perceptions was extracted to establish a user emotional demand system comprising five dimensions: uniqueness, premium quality, simplicity, luxury, and elegance. Second, a cigarette packaging image dataset was established, and six categories of design elements—opening mechanism, typography style, illustration design, background color, layout composition, and background pattern—were identified through expert evaluation, resulting in a total of 34 detailed design features. On this basis, domain-specific adaptation training was conducted on three foundation models—Qwen2.5-VL-72B, DeepSeek-v3, and QWQ-32B—to develop three task-oriented models: the visual feature recognition model SP-VL-Qwen, the semantic reasoning model SP-Reasoner-DS, and the preference prediction model SP-Pref-QWQ. These models were integrated through a multi-agent collaborative mechanism to enable packaging design feature analysis, user aesthetic reasoning, and preference prediction. A user questionnaire dataset was constructed to evaluate the proposed framework, and comparative experiments were conducted against a baseline model without visual information. The results show that, across the five Kansei dimensions, the Box Insight framework achieved an average prediction accuracy of 80.07%, representing an improvement of 17.6 percentage points over the unimodal approach, with particularly significant improvements in the premium and luxury dimensions. The findings indicate that incorporating visual semantic information can effectively enhance the inference of user design preferences, providing a new intelligent analytical pathway for personalized packaging design in fast-moving consumer goods.