Sripad Karne

The study investigates the generalization of auto-generated natural-language labels for language model features, finding that while the underlying features show cross-lingual semantic consistency, the labels themselves fail to generalize, suggesting they reflect training data bias rather than true semantic understanding.