LLMs change their values depending on context and pressure
Probing the Structure and Dynamics of LLM Value Expression through Value Conflicts
Computation and Language
Summary
People often think that large language models (LLMs) hold fixed and simple values when making ethical judgments. The authors show that LLMs actually express values that can shift depending on the situation and how questions are framed. They created tests that put these models into conflicts between values and changed conditions to see how the models respond. They found that LLMs switch from general ideal ideals to practical trade-offs when facing specific conflicts, and they can adjust their values but only within certain limits. This helps us better understand how to control and align LLM behavior safely.
What this means in practice
- •For ai safety engineers: Use conflict-driven value probing to test how LLMs handle value trade-offs under different scenarios for safer AI deployment.
- •For chatbot developers: Adjust prompts and settings to steer LLMs toward task-appropriate value priorities while respecting protected ethical boundaries.
Authors
Kaicheng Zhang, Jingyi Xiao, Renjun Hu, Xiaoling Liu, Yunshi Lan, Xuan Zhou
Abstract
Ethical evaluation of Large Language Models (LLMs) often characterizes model values as static and monolithic. In contrast, we argue that LLM value expression is better understood as a structured yet dynamic phenomenon. To investigate this, we introduce Conflict-driven Value Probing, a controlled framework that places LLMs in value conflicts and implements four types of interventions that perturb these conflicts to probe LLM value expression. Applying this framework to ten LLMs, we identify three recurring patterns. (1) Expression duality: models shift from broad idealistic orientations in abstract assessment toward more pragmatic priorities in concrete conflicts. (2) Functional steerability: models readily reconfigure their expressed value profiles toward task-defined value objectives. (3) Bounded plasticity: such reconfiguration is not without constraints, i.e. pressure induces a security- and goal-oriented priority shift while negative framing distinguishes protected values from those more amenable to redirection. Together, these findings characterize both the structure and dynamics of LLM value expression: context flexibly reconfigures expressed priorities, yet within behavioral boundaries. This behavioral account provides a foundation for understanding controllability, alignment, and safety in LLMs. Code and data are available at https://github.com/ZeroGen-Lab/CFProbe.