Relevance 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.

• Concreteness Filtering of SAE Latentsconcept0.846
  Pre-filtering step excluding abstract latents where off-topic detection is harder
• Sparse Autoencoders (SAE)method0.762
  Interpretability method criticized in this paper for shattering manifolds into atomic pieces, obscuring overarching semantic structure.
• SAEs can surface features relevant to meta-cognitive monitoring, not just object-level content representationclaim0.753
  Extension of mechanistic interpretability findings to the metacognitive domain
• SAE featuresconcept0.746
  The individual, supposedly monosemantic directions learned by SAEs; argued here to fragment manifolds into disconnected pieces.
• Textual SAE feature emotionality evaluationmethod0.745
  Method where Kimi evaluates steered vs unsteered text samples from another instance to rate SAE feature emotionality (0-100)
• Sparse Autoencoders (SAE) activation-based paradigmframework0.738
  Standard interpretability approach that VPD critiques and proposes an alternative to.
• Self-evaluated emotionality and textual evaluation of SAE features predict persistence in opposite directions.claim0.738
  Surprising finding that the two evaluation methods diverge in their relationship with persistence
• Single-Layer SAE Analysis Limitationconcept0.736
  Key limitation that prevents tracing inter-layer dynamics or how steering propagates through model depth