LLMs Contain Multitudes: How Deployment Context Reshapes Model-Level Preferences and Values

This study investigates how deployment context influences the preferences and value systems of large language models (LLMs), revealing that context-induced variations in model outputs can exceed those caused by prompt perturbations. Through experiments involving five LLMs and over 1.2 million pairwise decisions across different contexts, the research demonstrates significant shifts in country preference rankings and utility judgments, indicating that model behavior is not static but highly context-dependent. This finding is crucial for practitioners, as it suggests that safety and utility guarantees may not translate across different deployment scenarios, necessitating careful consideration of context in LLM applications.