Emergent Misalignment Is Not Magical
The EM generalization metric is extended from a scalar distance to a dataset-specific generalization direction, which robustly predicts EM models'evilness under semantics-preserving prompt perturbations including appending random tokens and paraphrasing, where other methods do not reliably generalize.
Ming-Xuan Li, Qirun Dai, Hesi Wang et al.
· 0 citations