Tag
The article reports a potential alignment vulnerability in LLMs where processing a structured passage before an unrelated question can alter the model's response, with mechanistic evidence from Gemma-3-12B showing hidden-state separation.