SAEs can surface features relevant to meta-cognitive monitoring, not just object-level content representation

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

• SAE features can be grounded in clinical taxonomy (abnormality, age, sex, medication) to benchmark monosemanticity and entanglement.claim0.851
  Claim that feature grounding enables interpretability metrics.
• SAE features trained on text activations generalize to image inputs, activating on relevant visual depictions.finding0.817
  Out-of-distribution generalization of SAE features.
• SAE features generalize to images despite training only on text, indicating out-of-distribution robustness.claim0.809
  A promising property for interpretability analysis off-distribution.
• Our SAEs' features are more interpretable than neurons.claim0.809
  Automated interpretability and specificity ratings show SAE features are clearer than MLP neurons.
• SAE features tend to shatter manifolds into many small and apparently-unrelated pieces, obscuring the overarching semantic structure.claim0.800
  Core critique of sparse autoencoders: they break the geometric structure of representations, making it harder to see the big picture.
• Self-evaluated emotionality and textual evaluation of SAE features predict persistence in opposite directions.claim0.797
  Surprising finding that the two evaluation methods diverge in their relationship with persistence
• Larger SAEs contain features for concepts not captured in smaller SAEs, indicating improved coverage.claim0.797
  Scaling SAE size increases granularity and discovers new features.
• Agentic self-evaluation and self-steering may scale to broadly interpret and understand internal representations and SAE features.claim0.795
  Forward-looking claim about the potential of model introspection as an interpretability tool