finding
active
finding:ablating-the-characters-remaining-subspace-affects-loss-only-on-newlines-and-patching-modulates-newline-predictionAblating the characters remaining subspace affects loss only on newlines, and patching modulates newline prediction
Causal validation.
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Central thesis.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Ablating the character count subspace has a large loss effect only when the next token is a newlinefinding0.893Causal validation.
- Linear separability.
- Author's mechanistic explanation for why regularization loss along persona directions is ineffective
- Specificity of illusion.
- Causal validation.
- Load-bearing interpretive claim about the layer-specificity of Burger et al.'s finding.
- Gradient-based attribution approximates ablation impact, enabling fast search for causally important features.
- Localizes truth representations to specific hidden states, motivating the rest of the analysis