finding
active
finding:summing-attribution-across-token-positions-reliably-highlights-effective-interpretable-features-unlike-max-or-per-position-displaySumming attribution across token positions reliably highlights effective interpretable features, unlike max or per-position display
Design choice validated empirically.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Papers (1)
paper
Methods (1)
method
- Backprop of logit-minus-mean-logit loss to surface causal intervention candidates, summed across token positions.
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.
- §5.5.
- A better basis (e.g., factoring tokens into languages or parts of speech) may yield a sparse interpretable modelhypothesis0.799Same model can be dense in one basis and sparse in another.
- Known limitation.
- Motivates nested SAE.
- Contrast in §6.3.
- Justifies the choice of response tokens for persona vector extraction in the pipeline
- Structure of the computational pathway.
- With raw data, wider token strides give more accurate approximations at same token budget.finding0.783Cost tradeoff.