Friends and Grandmothers in Silico: Localizing Entity Cells in Language Models

In three linesResearchers localize 'entity cells'—selective MLP neurons encoding entity-specific facts—across seven language models. On Qwen2.5-7B, suppressing a cell selectively erases recall for its matched entity while activating a single cell recovers knowledge even without context. These cells remain stable under aliases, acronyms, and multilingual forms.

Summary generated by Claude — human-verified