Can even finer-grained extraction and steering be achieved beyond the attention head level through integration of SAEs?

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

• SAE feature steering in history, conceptual, and zero-shot control conditions produces zero experience reports under either suppression or amplificationfinding0.781
  Shows gating effect is specific to the self-referential computational regime, not a general feature effect
• SAEs can surface features relevant to meta-cognitive monitoring, not just object-level content representationclaim0.780
  Extension of mechanistic interpretability findings to the metacognitive domain
• Steering base models along the assistant axis extracted from instruct models promotes helpful human archetypes and agreeable traits, suggesting the axis predates post-trainingfinding0.761
  Validation finding from Lu et al. 2026 supporting PSM's claim about pre-training persona structure
• SAE features trained on text activations generalize to image inputs, activating on relevant visual depictions.finding0.757
  Out-of-distribution generalization of SAE features.
• Inhibition steering produces larger accuracy drops than enhancement steering produces accuracy gains, across all models and datasets testedfinding0.756
  Key asymmetry finding: suppressing reflection is easier than inducing it.
• Steering vectors capture latent dimensions of reflective behavior more faithfully than surface-level embedding similarity.claim0.755
  Supported by the instruction discovery experiments comparing steering vs. embedding baselines.
• SOO fine-tuning could be extended to align AI representations of its own goals with human user preferences, reducing misalignment by fostering coherence between self-related and other-related preferenceshypothesis0.754
  Future work hypothesis about extending SOO to direct value alignment
• Steering vectors enable systematic discovery of reflection-inducing instructions beyond trial-and-error prompt design.claim0.754
  Core applied contribution claim, supported by top-k accuracy comparisons.