Reflection Inhibition via Activation Subtraction

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

• Reflection Enhancement via Activation Additionmethod0.819
  Adding steering vector in forward direction to push model activations toward stronger reflective behavior.
• The ease of suppressing reflection via activation steering raises security risks, as malicious actors could exploit reflection inhibition to bypass model safeguards.claim0.775
  Applied security implication derived from the asymmetry finding.
• Reflection is not merely a behavioral artifact of prompting but a phenomenon encoded in the model's activation space.claim0.768
  Central interpretive claim of the paper, supported by steering vector experiments.
• Contrastive pair activation subtractionmethod0.766
  Technique for obtaining concept vectors by presenting model with two scenarios differing in one respect and subtracting activations to isolate conceptual difference.
• Contrasting No Reflection with Triggered Reflection (µ(0→2)) provides a stronger reflection signal than contrasting Intrinsic with Triggered Reflection (µ(1→2)).claim0.747
  Empirical interpretation of which reference baseline yields more useful steering vectors.
• Suppressing reflection is considerably easier than inducing it, because inhibition requires the model to terminate reasoning while enhancement demands additional cognitive effort to re-examine reasoning trajectories.claim0.746
  Key asymmetry finding interpreted mechanistically by the authors.
• Activation steering of reflection has dual-use implications: it can enhance reflection as a defense mechanism, but malicious actors may also use it to inhibit reflection to facilitate jailbreaks.claim0.744
  Applied dual-use conclusion drawn from the paper's findings.
• Triggered Reflectionconcept0.732
  Reflection level where explicit cue words (e.g., 'wait') prompt the model to inspect and revise reasoning.