Auto-Interpretation of SAE Latents

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

• SAE Latentsconcept0.842
  Interpretable features extracted by sparse autoencoders used as steering targets in this study
• Relevance Filtering of SAE Latentsconcept0.827
  Pre-filtering step excluding latents naturally activated by each prompt to ensure genuine off-topic steering
• Sparse Autoencoders (SAE)method0.774
  Interpretability method criticized in this paper for shattering manifolds into atomic pieces, obscuring overarching semantic structure.
• SAE Latent Steeringmethod0.770
  Adding a multiple of the SAE latent decoder vector to token activations to causally test each latent's role in misalignment
• Concreteness Filtering of SAE Latentsconcept0.769
  Pre-filtering step excluding abstract latents where off-topic detection is harder
• Self-evaluated emotionality and textual evaluation of SAE features predict persistence in opposite directions.claim0.740
  Surprising finding that the two evaluation methods diverge in their relationship with persistence
• Sparse Autoencoders (SAE) activation-based paradigmframework0.735
  Standard interpretability approach that VPD critiques and proposes an alternative to.
• Textual SAE feature emotionality evaluationmethod0.734
  Method where Kimi evaluates steered vs unsteered text samples from another instance to rate SAE feature emotionality (0-100)