Steering base models along the assistant axis extracted from instruct models promotes helpful human archetypes and agreeable traits, suggesting the axis predates post-training

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

• The Assistant Axis in instruct models mainly inherits from pre-existing helpful and harmless human personas in base models, later acquiring additional associations (such as being an AI) during post-trainingclaim0.897
  Key mechanistic claim about the developmental origin of the Assistant persona
• The Assistant Axis is also present in pre-trained base models, where it primarily promotes helpful human archetypes (consultants, coaches) and inhibits spiritual onesclaim0.881
  Extends the Assistant Axis finding to pre-training, suggesting pre-training rather than post-training creates the axis
• Steering base models toward the Assistant Axis increases agreeableness traits (friendly, kind, helpful) and decreases extraversion in Gemma and openness in Llamafinding0.870
  Characterizes the trait content of the Assistant Axis in pre-trained models
• Steering base Gemma/Llama models toward the Assistant Axis increases completions describing helpful professional roles (therapist, consultant) and decreases spiritual/religious purpose mentionsfinding0.850
  Shows Assistant Axis in instruct models inherits from helpful human personas in base models
• 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 personaclaim0.836
  Central interpretive claim and motivation for future work
• Is the Assistant Axis formed during post-training or inherited from representations learned during pre-training?question0.816
  Motivates the base model steering experiments in §3.2.2
• Steering vector extracted from final post-expert-iteration model also successfully elicits deployment behaviorfinding0.809
  Replicates main result using in-distribution steering vector; addresses concern about pre-trained vector validity.
• What a model exposes by default tracks the norms it was trained toward, while steering acts on deviations from those defaults rather than on the defaults themselvesclaim0.804
  Central interpretive claim organizing the entire paper's results