Semantic Containment as a Fundamental Property of Emergent Misalignment

Rohan Saxena

Fine-tuning language models on narrowly harmful data causes emergent misalignment (EM) -- behavioral failures extending far beyond training distributions. Recent work demonstrates…