Sadism facet of Evil persona grows at the beginning of Evil persona development in OLMo-3 pretraining

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

• Sadism facet grows during Apertus pretraining, replicating OLMo-3 findingfinding0.880
  Replication of sadism growth finding on Apertus confirming cross-model generality
• Evil persona vector saturates by approximately 12B tokens in OLMo-3finding0.765
  Quantitative result showing Evil emerges earliest due to ubiquity and simplicity in pretraining data
• Humorous persona barely emerges by end of OLMo-3 pretraining, requiring pragmatic competencefinding0.759
  Quantitative result showing humor requires acquired linguistic competence and emerges latest
• Evil and Impolite persona vectors evolve in parallel in MDS space, suggesting intertwined representationsfinding0.752
  Geometric finding from MDS analysis suggesting shared representational structure
• 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 backfinding0.748
  Evidence that the evil persona region exhibits the stickiness hallmark of a genuine attractor basin
• Evil persona steering delta of +9.69 (p=0.011) at stage1-step3000 (12.6B tokens) in OLMo-3finding0.743
  First significant evil persona steering result; establishes emergence at 0.22% pretraining
• Description extraction emphasizes threatened egotism while Dialogue emphasizes idealism and reduces sadism in OLMo-3finding0.741
  Shows discourse-type-specific facet profiles supporting PSM hypothesis of diverse persona subforms
• Persona vectors extracted from base pretraining checkpoints steer fully post-trained OLMo-3-7B-Instructfinding0.740
  Shows persona directions persist through all alignment stages, answering key part of RQ2