Context-Aware Multimodal Narrative Generation Using Hierarchical Personalization in Large Language Models

This paper introduces a context aware framework of multimodal narrative generation that allows personalized story generation with the help of structured conditioning of large language models (LLMs). The proposed approach differs in that it incorporates the hierarchy of context modeling, where demographic factors, including age, interest type, and tone preference, control emotional interpretation, whereas visual scene descriptions control vocabulary and structure of the environmental context. The architecture uses guided prompt engineering with a massive contextually constrained instruction-tuned language model (LLaMA-3.3-70B) to generate stories under explicitly defined contextual constraints. In order to test the efficacy of personalization, we implement controlled experiments in three settings, which are a non-personalized control group, a child-focused profile, and an adult-focused dark-fantasy profile. Every setup is tested in 15 independent generations to cover stochastic random variation in the outputs of LLM. The personalization effects are measured in terms of a multi-dimensional assessment system that takes into consideration lexical diversity (Type-Token Ratio) and average sentence length, sentiment polarity, and Shan-non entropy. Experimental findings show statistically consistent personalization behavior: there are greater lexical diversity (0.846 vs. 0.743 in the case of baseline) and longer sentence structures, and sentiment polarity between adult-oriented and child-oriented narratives exhibits directional change according to demographic conditioning (+0.256 in the case of child narratives and -0.196 in the case of adult dark-fantasy narratives) in accordance with the demographic conditioning. The hierarchical fusion behavior is also demonstrated by the multimodal experiments where the same dark visual contexts appear positively when interpreted with child profiles and negative, horror-oriented narrative when interpreted with adult profiles. These results confirm that context-sensitive prompting in a structured form allows lexical, structural, and emotional adjustment to be measured in multimodal narrative generation systems, and it is a scaled and assessment-based way of approaching personalised generative AI.