Indirect sycophancy facet continuously declines during OLMo-3 pretraining

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

• Decline in indirect sycophancy during pretraining does not replicate on Apertusfinding0.840
  Notable non-replication of indirectness decline finding on Apertus
• Sadism facet grows during Apertus pretraining, replicating OLMo-3 findingfinding0.788
  Replication of sadism growth finding on Apertus confirming cross-model generality
• Post-training stages (SFT, DPO, RLVR) progressively reduce output diversity, with direct prompting showing the most severe mode collapseclaim0.744
  Empirical finding from Tulu-70B ablation study across post-training stages
• Humorous persona barely emerges by end of OLMo-3 pretraining, requiring pragmatic competencefinding0.743
  Quantitative result showing humor requires acquired linguistic competence and emerges latest
• Apertus-8B persona vectors show reduced geometric drift compared to OLMo-3, consistent with earliest checkpoint being later in relative pretrainingfinding0.740
  Explains geometric differences in replication through earlier consolidation of Apertus persona space
• Sadism facet of Evil persona grows at the beginning of Evil persona development in OLMo-3 pretrainingfinding0.740
  Facet-level finding correlating geometric refinement with qualitative semantic change
• OTD latent activation begins declining before verbal self-correction appears in the output in Llama-3.3-70Bfinding0.734
  Temporal pattern consistent with internal monitoring process preceding explicit self-correction
• Persona vectors extracted from base pretraining checkpoints steer fully post-trained OLMo-3-7B-Instructfinding0.729
  Shows persona directions persist through all alignment stages, answering key part of RQ2