method
active
method:logit-attribution-analysisLogit attribution analysis
Dot product of summed decoder vector with unembedding vector, following Hanna et al.
Neighborhood — ranked by edge-count
Papers (1)
paper
Methods (1)
method
- Adaptation of logit attribution to CLT features that decode into all subsequent layers.
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.
- Computing each feature's linear effect on output token logits via path expansion through MLP output weights and unembedding matrix
- The task of attributing model behaviors to specific training datapoints.
- Input-output mapping per feature across ~100 prompts.
- Measure causal importance of upstream variables.
- Cosine similarity at lags to test stationarity.
- Logit-minus-mean-logit loss was most effective for attribution, vs log-loss which was confusing to interpretfinding0.781Attribution loss selection.
- Samples from interpolated logits of original and abliterated models at α.
- Iterative sampling procedure recomputing before/after logits with each new token.