This research is dedicated to surmounting the limitations of Large Language Models (LLMs) in sentiment analysis. By delving into Myers-Briggs Type Indicator (MBTI) traits, we put forward an innovative Context-Aware Vectors method. Departing from traditional approaches such as rule-setting or data fine-tuning, we explore the latent interpretability of personality traits within LLMs. We conduct experiments on Meta-LLaMA-3.1-8B, employing the SCIFACT dataset devoid of personality bias and carefully crafted prompts. This enables us to extract context-aware vectors, by calculating the differences in the outputs of a specific layer under different MBTI Traits styles. When evaluated on the SemEval dataset, LLMs integrated with context-aware vectors, especially the INFJ-LLM, achieve state-of-the-art (SOTA) performance in comparison to 15 baseline models. Ablation experiments further verify the significance of each component of our context-aware vectors. In summary, this work offers a novel approach to integrating MBTI traits into LLMs, expands the scope of LLM interpretability research, and notably enhances LLMs’ performance in sentiment analysis.