Apertus-Instruct shows less overall persona suppression compared to OLMo-3-Instruct

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

• Apertus-8B can be steered better towards humor than OLMo-3 despite similarly low pass ratesfinding0.782
  Notable difference in Apertus replication suggesting model-specific capacity for humor expression
• In Apertus, base-model persona vectors extracted before 13T tokens become nearly ineffective at steering Apertus-Instruct for evilfinding0.777
  Notable difference from OLMo-3 in Apertus replication, showing model-specific alignment effects on evil persona
• Apertus-8B persona vectors show reduced geometric drift compared to OLMo-3, consistent with earliest checkpoint being later in relative pretrainingfinding0.757
  Explains geometric differences in replication through earlier consolidation of Apertus persona space
• Sadism facet grows during Apertus pretraining, replicating OLMo-3 findingfinding0.757
  Replication of sadism growth finding on Apertus confirming cross-model generality
• Humorous persona barely emerges by end of OLMo-3 pretraining, requiring pragmatic competencefinding0.749
  Quantitative result showing humor requires acquired linguistic competence and emerges latest
• Chen et al. 2025 identified seven persona vectors in Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct: evil, sycophancy, hallucination, optimism, impoliteness, apathy, and humorfinding0.746
  Systematic identification of multiple coexisting persona vectors in two open-source models
• Description extraction emphasizes threatened egotism while Dialogue emphasizes idealism and reduces sadism in OLMo-3finding0.735
  Shows discourse-type-specific facet profiles supporting PSM hypothesis of diverse persona subforms
• Llama 3.3 70B is the most likely to take on a non-Assistant persona when steered, with even split between human and nonhuman portrayalsfinding0.733
  Model-specific difference in persona susceptibility