This research explores strategies for steering the output of large language models (LLMs) towards specific styles, such as sentiment, emotion, or writing style, by adding style vectors to the activations of hidden layers during text generation. The study demonstrates that style vectors can be computed from recorded layer activations for input texts in a specific style, offering a simpler alternative to complex training-based approaches. Through experiments, the effectiveness of activation engineering using style vectors is shown to influence the style of generated text in a nuanced and parameterisable way, distinguishing it from prompt engineering. The research contributes to developing more adaptive and effective AI-empowered interactive systems. The paper introduces two approaches for generating style vectors: training-based style vectors and activation-based style vectors. Training-based style vectors are derived from generative steering vectors, while activation-based style vectors are obtained by aggregating layer activations for input sentences from the target style. The latter approach is more efficient and does not require additional optimization or prior knowledge about original styles. Experiments were conducted on datasets for sentiment, emotion, and writing style. The results show that activation-based style vectors perform well in capturing relevant style information and steering the model's output towards desired styles. The study also highlights the importance of choosing appropriate parameters, such as λ, to achieve effective style steering without producing nonsensical outputs. The research demonstrates that activation-based style vectors can be used to steer LLM outputs towards specific styles, such as sentiment, emotion, or writing style. The results indicate that these vectors provide smoother transitions and more controllable outputs compared to prompt engineering. The study also notes the potential for activation-based steering to generate new styles, expanding the possibilities beyond the constraints of pre-training knowledge. The paper concludes that activation-based style vectors offer a more efficient and effective approach for steering LLM outputs compared to training-based methods. The research also highlights the importance of considering ethical implications, such as the potential for generating biased or harmful content. The study emphasizes the need for further research to address these challenges and improve the overall performance and fairness of AI systems.This research explores strategies for steering the output of large language models (LLMs) towards specific styles, such as sentiment, emotion, or writing style, by adding style vectors to the activations of hidden layers during text generation. The study demonstrates that style vectors can be computed from recorded layer activations for input texts in a specific style, offering a simpler alternative to complex training-based approaches. Through experiments, the effectiveness of activation engineering using style vectors is shown to influence the style of generated text in a nuanced and parameterisable way, distinguishing it from prompt engineering. The research contributes to developing more adaptive and effective AI-empowered interactive systems. The paper introduces two approaches for generating style vectors: training-based style vectors and activation-based style vectors. Training-based style vectors are derived from generative steering vectors, while activation-based style vectors are obtained by aggregating layer activations for input sentences from the target style. The latter approach is more efficient and does not require additional optimization or prior knowledge about original styles. Experiments were conducted on datasets for sentiment, emotion, and writing style. The results show that activation-based style vectors perform well in capturing relevant style information and steering the model's output towards desired styles. The study also highlights the importance of choosing appropriate parameters, such as λ, to achieve effective style steering without producing nonsensical outputs. The research demonstrates that activation-based style vectors can be used to steer LLM outputs towards specific styles, such as sentiment, emotion, or writing style. The results indicate that these vectors provide smoother transitions and more controllable outputs compared to prompt engineering. The study also notes the potential for activation-based steering to generate new styles, expanding the possibilities beyond the constraints of pre-training knowledge. The paper concludes that activation-based style vectors offer a more efficient and effective approach for steering LLM outputs compared to training-based methods. The research also highlights the importance of considering ethical implications, such as the potential for generating biased or harmful content. The study emphasizes the need for further research to address these challenges and improve the overall performance and fairness of AI systems.