Technology & Tools

LLMs correctly identify human values just 68.3% of the time

Testing 21 instruction-tuned LLMs against Schwartz's ten basic values, researchers found a pooled first-choice accuracy of 68.3% across 1,000 labeled texts.

Why it mattersLLMs are increasingly used in ethics, moderation, and alignment work, so missing the intended value on the first try nearly a third of the time is a concrete reliability gap.