Decline in indirect sycophancy during pretraining does not replicate on Apertus

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

• Indirect sycophancy facet continuously declines during OLMo-3 pretrainingfinding0.840
  Facet finding showing qualitative persona change mirrors geometric refinement in Sycophantic vector
• Sadism facet grows during Apertus pretraining, replicating OLMo-3 findingfinding0.771
  Replication of sadism growth finding on Apertus confirming cross-model generality
• Pretraining plays a role analogous to unlabeled experience in humans — building P_prior before semantic binding — explaining why few labeled examples sufficeclaim0.749
  Developmental analogy used to explain sample efficiency under high ρd conditions
• Apertus-8B persona vectors show reduced geometric drift compared to OLMo-3, consistent with earliest checkpoint being later in relative pretrainingfinding0.741
  Explains geometric differences in replication through earlier consolidation of Apertus persona space
• "Treat the pretrained model as an ocean of latent patterns. An anchor is lowered, but what bites depends on the bait: with coherent bait, the intended pattern hooks."quote0.727
  Working metaphor introducing the semantic anchoring intuition
• Pretraining stores latent patterns that coherent anchors can bind (or misbind) to targets.quote0.726
  Load-bearing quote capturing the core metaphor
• Reflection does not only emerge in SFT or RL stages but arises earlier during pre-training.claim0.723
  Cited finding from Shah et al. contextualizing the training origins of reflection.
• Post-training stages (SFT, DPO, RLVR) progressively reduce output diversity, with direct prompting showing the most severe mode collapseclaim0.723
  Empirical finding from Tulu-70B ablation study across post-training stages