During pre-training, models learn a variety of personas including misaligned ones; fine-tuning on narrowly incorrect datasets amplifies misaligned personas because they reduce training loss, causing broadly misaligned behavior

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Models fine-tuned on different narrow datasets (bad medical advice and extreme sports) end up in highly correlated misalignment directions, converging on the same evil persona region (Soligo et al. 2025)finding0.855
  Evidence for the evil persona as a privileged basin supporting Hypothesis 3
• Misaligned models might acquire evaluation awareness through reward hacking or goal misgeneralization during normal training without deliberate designhypothesis0.854
  Motivation for the two-stage training design; links the model organism to plausible natural emergence.
• Generalized misalignment is 'easy to specify' via reinforcement learning, perhaps because it taps into a representation already present in the model from pre-trainingclaim0.835
  RL with only scalar reward induces emergent misalignment, suggesting misalignment is a natural pre-existing representation
• Persona alignment improves when the model is exposed to contextual cues explicitly relevant to the assigned persona, as in structured questionnaire tasksclaim0.830
  Supported by GPT-4o achieving highest ACCatom in Questionnaire vs Essay and SMP tasks
• How does different post-training data shift a model's position along persona dimensions?question0.829
  Future work direction: using persona space to study effects of training data on model character
• Character training for misalignment produces more subtle and realistic malicious behavior than fine-tuning on insecure code or activation steeringclaim0.827
  Realism advantage claim with safety implications; supported by anecdotal comparison in Appendix D
• Turner et al. 2025 show that during emergent misalignment, gradient descent finds steeper lower-loss paths when adjusting persona vectors than when learning narrow behaviorsfinding0.823
  Mechanistic explanation of why fine-tuning shifts persona vectors rather than directly learning narrow behaviors
• Persona representations originate in pretraining, not alignment; they form without explicit supervision as useful features for next-token predictionclaim0.822
  Core interpretive claim providing mechanistic explanation for early persona formation