llm-preferences

This paper investigates whether large language models have stable preferences across different deployment contexts, finding that context can cause larger variations than prompt perturbations, suggesting that measured preferences are context-conditioned rather than fixed properties.