LLM Value Recognition Errors via Schwartz Theory Framework
July 23, 2026
An evaluation of 21 instruction-tuned models reveals an Acc@1 of 0.683 when identifying Schwartz's ten basic values in situational contexts. Errors are frequently concentrated between adjacent values, suggesting models struggle with fine-grained semantic distinctions in value-based reasoning.
HOW THIS AFFECTS YOU
●
researcherThis study highlights the need for more granular evaluation frameworks for model alignment and value recognition.
●
policyExpect inconsistencies in how models interpret and apply social or ethical values in nuanced scenarios.