Concreteness Filtering of SAE Latents

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

• Relevance Filtering of SAE Latentsconcept0.846
  Pre-filtering step excluding latents naturally activated by each prompt to ensure genuine off-topic steering
• SAE Latentsconcept0.790
  Interpretable features extracted by sparse autoencoders used as steering targets in this study
• SAE features can be grounded in clinical taxonomy (abnormality, age, sex, medication) to benchmark monosemanticity and entanglement.claim0.741
  Claim that feature grounding enables interpretability metrics.
• SAEs can surface features relevant to meta-cognitive monitoring, not just object-level content representationclaim0.738
  Extension of mechanistic interpretability findings to the metacognitive domain
• Sparse Autoencoders (SAE)method0.736
  Interpretability method criticized in this paper for shattering manifolds into atomic pieces, obscuring overarching semantic structure.
• Self-evaluated emotionality and textual evaluation of SAE features predict persistence in opposite directions.claim0.723
  Surprising finding that the two evaluation methods diverge in their relationship with persistence
• Concreteness Judgemethod0.722
  LLM-based judge rating SAE latent labels 0-100 for concreteness to filter steering candidates
• Feature attribution via gradient dot product with SAE decodermethod0.718
  Computing attribution as the dot product of the output logit gradient with the SAE decoder weight, multiplied by feature activation.