Response-average token extraction yields stronger steering effectiveness than prompt-last or prompt-average extraction positions

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

• Response-Average Token Extractionmethod0.817
  Strategy of extracting persona vectors from averaged activations over response tokens, found most effective compared to prompt-based positions
• Final token position consistently yields the strongest truth interventions across modelsfinding0.802
  Experiment 1 finding on token position, consistent with prior work
• Steering vector extracted from final post-expert-iteration model also successfully elicits deployment behaviorfinding0.797
  Replicates main result using in-distribution steering vector; addresses concern about pre-trained vector validity.
• Stepwise steering achieves over 5% accuracy improvement compared to all-token intervention at similar token budgetfinding0.790
  Key result demonstrating advantage of stepwise over all-token steering strategy
• Stepwise steering preserves accuracy while reducing cost, whereas all-token steering causes significant degradation at large intervention strengthsclaim0.785
  Comparative claim between the two steering strategies
• Activation steering elicits deployment behavior throughout all four rounds of expert iterationfinding0.778
  Shows steering remains effective even as model becomes more robust to prompting-based attempts to elicit deployment behavior.
• How can we be sure that steering methods actually elicited the deployment behavior, as opposed to only suppressing verbalizations of being deployed?question0.776
  Central motivating question of the paper; the model organism approach is the proposed answer.
• Activation steering is not equivalent to simply adding the deployment-side contrastive prompts to the user messageclaim0.775
  Key distinction showing steering offers value beyond prompting; supported by Figure 5 and random vector experiments.