SAE features trained on text activations generalize to image inputs, activating on relevant visual depictions.

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

• SAE features generalize to images despite training only on text, indicating out-of-distribution robustness.claim0.888
  A promising property for interpretability analysis off-distribution.
• Sparse Autoencoders (SAE) activation-based paradigmframework0.828
  Standard interpretability approach that VPD critiques and proposes an alternative to.
• SAEs can surface features relevant to meta-cognitive monitoring, not just object-level content representationclaim0.817
  Extension of mechanistic interpretability findings to the metacognitive domain
• Feature attribution via gradient dot product with SAE decodermethod0.796
  Computing attribution as the dot product of the output logit gradient with the SAE decoder weight, multiplied by feature activation.
• SAE features can be grounded in clinical taxonomy (abnormality, age, sex, medication) to benchmark monosemanticity and entanglement.claim0.793
  Claim that feature grounding enables interpretability metrics.
• Self-evaluated emotionality and textual evaluation of SAE features predict persistence in opposite directions.claim0.790
  Surprising finding that the two evaluation methods diverge in their relationship with persistence
• Sparse Autoencoders (SAE)method0.789
  Interpretability method criticized in this paper for shattering manifolds into atomic pieces, obscuring overarching semantic structure.
• Automated interpretability (Claude 3 Opus) and specificity scoring show SAE features are significantly more interpretable and specific than MLP neurons.finding0.782
  Quantitative comparison supporting SAE utility.