finding
active
finding:ablating-the-character-count-subspace-has-a-large-loss-effect-only-when-the-next-token-is-a-newlineAblating the character count subspace has a large loss effect only when the next token is a newline
Causal validation.
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Models represent scalar quantities on low-dimensional rippled feature manifolds that optimally trade off capacity against distinguishabilityassociated_withsupportsCentral 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.
- Causal validation.
- Linear separability.
- Specificity of illusion.
- Nonspecific to self.
- Author's mechanistic explanation for why regularization loss along persona directions is ineffective
- Feature attribution (gradient-based) correlates 0.8 with ablation effects on the 'John' and 'Kobe' examples.finding0.736Validation of attribution as a fast proxy for causal importance.
- Ablation depends on t_entity.
- Central practical contribution: the CL loss offers a viable mitigation strategy