Features can be used to steer large models.

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

• will these methods work for large models?question0.783
  Motivating question for the paper, addressed by scaling SAEs to Claude 3 Sonnet.
• 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.
• Whether the steerability map transfers across models and survives fine-tuning serves as the future research avenuehypothesis0.770
  Identified as the primary open question at the end of the paper
• Features may not be strictly one-dimensional objects; higher-dimensional feature manifolds may exist in model representationshypothesis0.768
  Extension of superposition hypothesis to account for continuous families of features
• 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 themselvesclaim0.768
  Central interpretive claim organizing the entire paper's results
• Larger SAEs contain features for concepts not captured in smaller SAEs, indicating improved coverage.claim0.767
  Scaling SAE size increases granularity and discovers new features.
• Base models are good modellers of worlds but not of their own state, because they lack a developed self-model initially.claim0.763
  Observation about asymmetry in base model capabilities.
• Optimally steering model behavior requires isolating concept geometry and defining operators to navigate it.claim0.762