This paper presents a comprehensive study of concept guidance in large language models (LLMs), focusing on the ability to control the behavior of LLMs by manipulating their hidden representations. The authors extend previous work on truthfulness to a broader set of concepts, including appropriateness, humor, creativity, and quality. They introduce a novel metric, perplexity-normalized effect size (PNES), to evaluate both the success of concept elicitation and the fluency of guided models. The study shows that while some concepts like truthfulness are relatively easy to guide, others such as appropriateness remain challenging to elicit and may lead to confusion. The authors also find that probes with high detection accuracy do not necessarily lead to optimal guidance, contradicting previous findings for truthfulness. This suggests that detectability and guidability are not directly correlated, and that the nature of the concept plays a significant role in its guidability. The paper evaluates various detection and guidance techniques, including logistic regression, difference-in-means, and principal component analysis. It demonstrates that logistic regression performs best in terms of detection accuracy for late layers in Llama-2-chat and Mistral-instruct. The study also highlights the trade-off between concept elicitation and fluency degradation, showing that strong guidance can lead to 'gibberish' outputs. The authors propose a framework for concept guidance that allows for the manipulation of hidden representations along linear directions, enabling targeted control over model behavior. They demonstrate that this approach can be used to guide models along multiple directions simultaneously, such as truthfulness and compliance. The results show that while some concepts are more easily guided, others require careful tuning and may not be consistently effective across all models and settings. The study underscores the complexity of concept guidance in LLMs and the need for further research into the interplay between detectability, guidability, and the nature of the concept. The authors hope that their work will inspire stronger follow-up approaches and contribute to the development of more robust techniques for concept guidance in LLMs.This paper presents a comprehensive study of concept guidance in large language models (LLMs), focusing on the ability to control the behavior of LLMs by manipulating their hidden representations. The authors extend previous work on truthfulness to a broader set of concepts, including appropriateness, humor, creativity, and quality. They introduce a novel metric, perplexity-normalized effect size (PNES), to evaluate both the success of concept elicitation and the fluency of guided models. The study shows that while some concepts like truthfulness are relatively easy to guide, others such as appropriateness remain challenging to elicit and may lead to confusion. The authors also find that probes with high detection accuracy do not necessarily lead to optimal guidance, contradicting previous findings for truthfulness. This suggests that detectability and guidability are not directly correlated, and that the nature of the concept plays a significant role in its guidability. The paper evaluates various detection and guidance techniques, including logistic regression, difference-in-means, and principal component analysis. It demonstrates that logistic regression performs best in terms of detection accuracy for late layers in Llama-2-chat and Mistral-instruct. The study also highlights the trade-off between concept elicitation and fluency degradation, showing that strong guidance can lead to 'gibberish' outputs. The authors propose a framework for concept guidance that allows for the manipulation of hidden representations along linear directions, enabling targeted control over model behavior. They demonstrate that this approach can be used to guide models along multiple directions simultaneously, such as truthfulness and compliance. The results show that while some concepts are more easily guided, others require careful tuning and may not be consistently effective across all models and settings. The study underscores the complexity of concept guidance in LLMs and the need for further research into the interplay between detectability, guidability, and the nature of the concept. The authors hope that their work will inspire stronger follow-up approaches and contribute to the development of more robust techniques for concept guidance in LLMs.