Feature attribution via gradient dot product with SAE decoder

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

• Gradient-based data attributionmethod0.800
  Baseline method against which probe-based ranking is compared; more computationally expensive.
• SAE features generalize to images despite training only on text, indicating out-of-distribution robustness.claim0.798
  A promising property for interpretability analysis off-distribution.
• SAE features trained on text activations generalize to image inputs, activating on relevant visual depictions.finding0.796
  Out-of-distribution generalization of SAE features.
• Sparse Autoencoders (SAE)method0.774
  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 representationclaim0.762
  Extension of mechanistic interpretability findings to the metacognitive domain
• TopK Sparse Autoencoders (SAEs)method0.757
  Sparse dictionary learning method used to extract interpretable features from EEG transformer embeddings.
• SAE training loss (MSE + L1 penalty with decoder norm scaling)method0.755
  The objective function combining L2 reconstruction error and L1 penalty scaled by decoder norm, used to train the SAE.
• Feature attribution (gradient-based) correlates 0.8 with ablation effects on the 'John' and 'Kobe' examples.finding0.753
  Validation of attribution as a fast proxy for causal importance.