Pairwise steering needs empirical maps rather than assuming vector addition will behave like ordinary semantic blending

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

• Steering vector constructed from all 16 contrastive pairs outperforms most single-pair vectors; best 4-pair vector outperforms full 16-pair vectorfinding0.788
  Demonstrates averaging multiple prompt pairs reduces noise; optimal subset selection further improves performance.
• Steering vectors capture latent dimensions of reflective behavior more faithfully than surface-level embedding similarity.claim0.777
  Supported by the instruction discovery experiments comparing steering vs. embedding baselines.
• Linear steering is often mismatched with a model's internal representation geometry, producing noisy, off-target effects.claim0.777
  The paper's critique of the standard linear steering baseline, supported by the days-of-week demo.
• Our method enables bidirectional steering of model behavior.finding0.776
  The method can steer the model in both positive and negative directions on the target semantic.
• Steering vectors discover effective triggers such as 'However' and 'Otherwise', consistent with prior reported reflection datasetsfinding0.770
  Validates that steering vectors capture reflection semantics by finding tokens reported in related work.
• Steering preferentially amplifies exaggerated and undesirable styles, not competence-oriented or assistant-default behaviorsclaim0.769
  The top of the steerability ranking is dominated by exaggerated or attention-grabbing styles
• Optimally steering model behavior requires isolating concept geometry and defining operators to navigate it.claim0.769
• Whether the steerability map transfers across models and survives fine-tuning serves as the future research avenuehypothesis0.766
  Identified as the primary open question at the end of the paper