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

In three linesResearchers localize 'entity cells' in MLP neurons across language models (Qwen2.5-7B, etc.). These selectively activated neurons encode entity-specific facts. Suppressing one cell erases recall for that entity alone; activating it recovers knowledge even without context. Cells remain stable across aliases, acronyms, and multilingual forms.

Summary generated by Claude — human-verified