Post-training elicits rather than installs personas; the persona direction is already laid down in pretraining

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

• post-training steers models toward a particular region of persona space but only loosely tethers them to itquote0.879
  Load-bearing summary of the paper's core finding about persona stability
• Persona directions progressively refine during pretraining with most refinement happening early, stabilizing at later checkpointsfinding0.877
  Answers RQ2 geometrically: adjacent-checkpoint cosine similarity stays high but step-to-step movement is largest early
• Post-training steers models toward a particular region of persona space but only loosely tethers them to it, motivating work on training and steering strategies that more deeply anchor models to a coherent personaclaim0.867
  Central interpretive claim and motivation for future work
• Persona representations originate in pretraining, not alignment; they form without explicit supervision as useful features for next-token predictionclaim0.838
  Core interpretive claim providing mechanistic explanation for early persona formation
• How does different post-training data shift a model's position along persona dimensions?question0.835
  Future work direction: using persona space to study effects of training data on model character
• Post-training is key to eliciting strong introspective awareness; base pretrained models do not show above-chance detectionclaim0.823
  Finding that base models have high false positives and no net positive performance.
• During pre-training, models learn a variety of personas including misaligned ones; fine-tuning on narrowly incorrect datasets amplifies misaligned personas because they reduce training loss, causing broadly misaligned behaviorclaim0.820
  The paper's central mechanistic explanation of why narrow fine-tuning causes broad misalignment
• Can pretraining-stage interventions through data filtering, gradient-based methods, or architectural modifications target persona directions as they form?question0.818
  Practical safety question about intervening on persona representations during pretraining