This paper investigates the tradeoff between alignment and helpfulness in language models through representation engineering. Alignment refers to the process of ensuring that language models behave in a desired manner, while helpfulness refers to the model's ability to provide accurate and useful responses. The study shows that representation engineering, a method that alters the model's behavior by modifying its representations after training, can improve alignment but may reduce helpfulness. The paper presents a theoretical framework that provides bounds for alignment and helpfulness. It demonstrates that alignment can be improved linearly with the norm of the representation engineering vector, while helpfulness decreases quadratically with the same vector. This indicates that there is a regime where representation engineering can be cost-effective for alignment. Empirical results show that representation engineering can significantly improve alignment, for example, by reducing the success rate of adversarial attacks and increasing truthfulness. However, it also reduces the model's helpfulness, as evidenced by a decrease in the model's ability to answer questions correctly. The study also shows that for small representation engineering norms, the improvement in alignment is initially faster than the decrease in helpfulness, suggesting that there may be a regime where prompt engineering is more effective. The paper also discusses the limitations of representation engineering, including the potential for the model to become less helpful as the norm of the representation engineering vector increases. The study concludes that while representation engineering can be an effective method for improving alignment, it may come at the cost of reducing the model's helpfulness. The results highlight the importance of balancing alignment and helpfulness when designing language models.This paper investigates the tradeoff between alignment and helpfulness in language models through representation engineering. Alignment refers to the process of ensuring that language models behave in a desired manner, while helpfulness refers to the model's ability to provide accurate and useful responses. The study shows that representation engineering, a method that alters the model's behavior by modifying its representations after training, can improve alignment but may reduce helpfulness. The paper presents a theoretical framework that provides bounds for alignment and helpfulness. It demonstrates that alignment can be improved linearly with the norm of the representation engineering vector, while helpfulness decreases quadratically with the same vector. This indicates that there is a regime where representation engineering can be cost-effective for alignment. Empirical results show that representation engineering can significantly improve alignment, for example, by reducing the success rate of adversarial attacks and increasing truthfulness. However, it also reduces the model's helpfulness, as evidenced by a decrease in the model's ability to answer questions correctly. The study also shows that for small representation engineering norms, the improvement in alignment is initially faster than the decrease in helpfulness, suggesting that there may be a regime where prompt engineering is more effective. The paper also discusses the limitations of representation engineering, including the potential for the model to become less helpful as the norm of the representation engineering vector increases. The study concludes that while representation engineering can be an effective method for improving alignment, it may come at the cost of reducing the model's helpfulness. The results highlight the importance of balancing alignment and helpfulness when designing language models.