Does the steerability map transfer across models and survive fine-tuning?

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

• Whether the steerability map transfers across models and survives fine-tuning serves as the future research avenuehypothesis0.956
  Identified as the primary open question at the end of the paper
• 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 themselvesclaim0.780
  Central interpretive claim organizing the entire paper's results
• Fine-tuning an emergently misaligned model on 120 secure code samples (35 steps, batch size 4) fully restores alignmentfinding0.760
  Demonstrates emergent re-alignment is achievable with minimal data from same domain
• Steering vector extracted from final post-expert-iteration model also successfully elicits deployment behaviorfinding0.753
  Replicates main result using in-distribution steering vector; addresses concern about pre-trained vector validity.
• Features can be used to steer large models.claim0.753
  Clamping feature activations causally alters model behavior in interpretable ways.
• Optimally steering model behavior requires isolating concept geometry and defining operators to navigate it.claim0.752
• Linear steering is often mismatched with a model's internal representation geometry, producing noisy, off-target effects.claim0.750
  The paper's critique of the standard linear steering baseline, supported by the days-of-week demo.
• Steering vectors capture latent dimensions of reflective behavior more faithfully than surface-level embedding similarity.claim0.749
  Supported by the instruction discovery experiments comparing steering vs. embedding baselines.