Base-model persona vector analyses carry through to post-trained descendants, validating safety audits on base checkpoints

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

• Persona vectors extracted from base pretraining checkpoints steer fully post-trained OLMo-3-7B-Instructfinding0.846
  Shows persona directions persist through all alignment stages, answering key part of RQ2
• Persona vector-based data filtering and LLM judge-based data filtering have complementary strengths for identifying problematic training dataclaim0.804
  Author's interpretive conclusion from comparing filtering strategies
• Persona vectors are present from the earliest available Apertus-8B checkpoint at 1.4% of pretraining (210B tokens)finding0.795
  Replication finding on Apertus confirming early persona formation generalizes across model families
• There may exist a natural 'persona basis' characterizing the space of all model personas, with correlations between persona vectors predicting co-expression of traitshypothesis0.789
  Open question proposed by authors for future work on the dimensionality and structure of persona space
• Persona vectors capture signal specific to their assigned trait, as evidenced by within-trait correlations exceeding cross-trait baselinesclaim0.782
  Author's interpretation establishing that persona vectors are not merely general misalignment indicators
• Single-method persona safety evaluation is incomplete: prompting and activation steering expose different, architecture-dependent vulnerability profiles, and testing with only one method can miss a model's dominant failure mode.claim0.782
  Central thesis of the paper
• In Apertus, base-model persona vectors extracted before 13T tokens become nearly ineffective at steering Apertus-Instruct for evilfinding0.778
  Notable difference from OLMo-3 in Apertus replication, showing model-specific alignment effects on evil persona
• We hypothesize that axes of persona differentiation within LLMs are likely already present in base models and inherited from the pre-training corpushypothesis0.777
  Motivated by near-identical PCs for base and instruct Gemma