Tag
Researchers trace how LLMs recall relational facts by probing per-head attention contributions, showing these are strong linear features whose fidelity correlates with relation specificity and entity connectedness.