realignment

This paper investigates the robustness of emergent misalignment in language models, finding that both misalignment and realignment are highly sensitive to superficial dataset characteristics and that previously reported mechanistic signatures do not consistently correlate with behavioral changes.