Dual Mechanisms of Value Expression: Intrinsic vs. Prompted Values in Large Language Models

Large language models can express values in two main ways: (1) $\textit{intrinsic}$ expression, reflecting the model's inherent values learned during training, and (2) $\textit{prompted}$ expression, elicited by explicit prompts. Given their widespread use in value alignment, it is paramount to clearly understand their underlying mechanisms, particularly whether they mostly overlap (as one might expect) or rely on distinct mechanisms. We analyze this largely understudied problem at the mechanistic level using two approaches: (1) $\textit{value vectors}$, feature directions representing value mechanisms extracted from the residual stream, and (2) $\textit{value neurons}$, MLP neurons that contribute to value vectors. We demonstrate that intrinsic and prompted value mechanisms partly share common components crucial for inducing value expression, generalizing across languages and reconstructing theoretical inter-value correlations in the model's internal representations. Yet, each mechanism also possesses unique components that fulfill distinct roles. In particular, the intrinsic mechanism activates in more diverse value-related scenarios and promotes response diversity, whereas the prompted mechanism strengthens instruction compliance, taking effect even in distant tasks like jailbreaking.

As AI assistants become widely used, they are increasingly expected to reflect human values (e.g., caring for others, striving for achievement, respecting tradition...). We can make models express these values using two different methods: either by training them to answer that way, or by inserting an explicit prompt prompt to express a value. However, while these two methods are often treated as to be interchangeable, it is not really clear whether they work in the same way inside the model. We study this question by looking at the internal activity of language models while they express different human values. We find that the two types of value expression partly share a common mechanism: both capture meaningful value concepts and relationships among values. However, they also differ in important ways. Intrinsic values are associated with mechanisms that support more varied and natural responses, whereas prompted values are more strongly linked to mechanisms that make the model follow instructions. This distinction matters because training and prompting are two common ways to shape AI behavior. Our results suggest that they are not simple substitutes: both have their own strengths, and understanding how these two mechanisms are similar and different can help people choose the right approach for different situations.