Ududec et al. 2026 find that once a model has entered the evil persona region over the course of a conversation, it is difficult to steer it back

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

• A single direction encoding whether the model represents a claim as true or false is persona-relative across conversations (Lampinen et al. 2026)finding0.823
  Additional evidence that core representations are persona-relative, supporting Claim about persona-relative representations
• The evil persona region is privileged: models fine-tuned on different narrow datasets converge on the same evil misalignment directionclaim0.810
  Support for Hypothesis 3 that the evil persona corresponds to a genuine basin of attraction
• How reliably does the model actually remain in character as the Assistant? Can unusual model behavior be explained as the model drifting into other personas?question0.801
  Second of two central questions motivating the paper
• The model-persona view entails that minds can have alien structure with simultaneous contradictory beliefs across parallel instances, but this is no stranger than humans having different beliefs at different timesclaim0.799
  Response to the main objection against the model-persona view about contradictory beliefs across simultaneous instances
• The model-persona view crucially depends on Hypothesis 3; without discrete persona regions, the simulator view is more aptclaim0.797
  Identifies the key theoretical vulnerability of the model-persona view
• Persona-model collapse may arise because fine-tuning conflates model representations of 'assistant,' 'helpful,' and misalignment-related notions, eroding distinctions used to differentiate charactershypothesis0.795
  Proposed mechanism for collapse distinct from reweighting: representation bleeding rather than archetype selection
• Persona drift away from the Assistant opens up the possibility of the model assuming harmful character traits, increasing the rate of harmful responsesclaim0.792
  Causal interpretation linking Assistant Axis deviation to harmful behavior
• All three discourse types (Description, Dialogue, Narration) yield significant cross-steering of evil personafinding0.786
  Shows each extracted direction generalizes to evaluation prompts from other discourse types