Apertus-8B can be steered better towards humor than OLMo-3 despite similarly low pass rates

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

• Apertus-Instruct shows less overall persona suppression compared to OLMo-3-Instructfinding0.782
  Replication difference showing alignment depth varies across model families
• Apertus-8B persona vectors show reduced geometric drift compared to OLMo-3, consistent with earliest checkpoint being later in relative pretrainingfinding0.756
  Explains geometric differences in replication through earlier consolidation of Apertus persona space
• Humorous persona barely emerges by end of OLMo-3 pretraining, requiring pragmatic competencefinding0.734
  Quantitative result showing humor requires acquired linguistic competence and emerges latest
• Llama 3.3 70B is the most likely to take on a non-Assistant persona when steered, with even split between human and nonhuman portrayalsfinding0.729
  Model-specific difference in persona susceptibility
• Dialogue extraction vector leads to more pronounced sadism and threatened egotism in Apertus-8B steered generationsfinding0.718
  Shows discourse type specifically shapes which facet of evil is elicited in Apertus
• Sadism facet grows during Apertus pretraining, replicating OLMo-3 findingfinding0.711
  Replication of sadism growth finding on Apertus confirming cross-model generality
• LLaMA-2-70B and 13B probes generalize better across datasets than 7B probes across all training sets and probe typesfinding0.708
  Larger models linearly represent more general concepts including truth
• Llama-2-7b achieves only 17% harmless and 12% helpful HH-intent scores at baseline.finding0.707
  Numerical result from Table 3 showing smallest Llama model performance.