Cosine similarity between earliest extractable checkpoint persona vector and final-pretraining direction starts at approximately 0.3

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

• Apertus-8B persona vectors show reduced geometric drift compared to OLMo-3, consistent with earliest checkpoint being later in relative pretrainingfinding0.847
  Explains geometric differences in replication through earlier consolidation of Apertus persona space
• Persona vectors extracted from base pretraining checkpoints steer fully post-trained OLMo-3-7B-Instructfinding0.845
  Shows persona directions persist through all alignment stages, answering key part of RQ2
• Persona vectors are present from the earliest available Apertus-8B checkpoint at 1.4% of pretraining (210B tokens)finding0.820
  Replication finding on Apertus confirming early persona formation generalizes across model families
• Cosine Similarity of Persona Vectorsconcept0.816
  Geometric measure used to track persona vector drift and stability across checkpoints
• Same-trait cosine similarities between GPT-4.1-mini and DeepSeek-V4-Flash extracted persona vectors are ≥0.93finding0.809
  Validates robustness of persona directions to choice of LLM judge model
• Differences in persona vector geometry across checkpoints correspond to meaningfully different versions of the personaclaim0.809
  Connects geometric refinement findings to semantic/facet changes in persona expression
• Persona directions progressively refine during pretraining with most refinement happening early, stabilizing at later checkpointsfinding0.809
  Answers RQ2 geometrically: adjacent-checkpoint cosine similarity stays high but step-to-step movement is largest early
• Persona vectors form within 0.22% of OLMo-3 pretraining (≈12.6B tokens)finding0.800
  Core quantitative result answering RQ1 for OLMo-3