method
active
method:logit-weight-analysisLogit Weight Analysis
Computing each feature's linear effect on output token logits via path expansion through MLP output weights and unembedding matrix
Neighborhood — ranked by edge-count
Frameworks (1)
framework
- Prior Anthropic paper enabling circuit-level analysis of attention-only transformers; motivates current MLP decomposition
Concepts (1)
concept
- Anomalous Tokensassociated_withExtremely rare or never-used vocabulary elements that may distort logit weight analysis; excluded from feature analysis
Claims (1)
claim
- Authors argue features are model properties because logit effects and ablations are consistent with feature interpretations
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.
- Dot product of summed decoder vector with unembedding vector, following Hanna et al.
- Correlating logit weight vectors between features from different models as a measure of downstream-effect universality
- Samples from interpolated logits of original and abliterated models at α.
- Training-free monitor that token-forces an eval-aware sentence and computes its probability at each CoT sentence boundary.
- Primary self-report measure: probability-weighted expected value over all ten digit-token logits, yielding a continuous rating that preserves full distributional signal
- The paper's method: interpolate between two models in logit space, measure compliance where common, and extrapolate to the original.
- Parameter-free loss transformation applied to each task loss to equalize scales
- Prior method extrapolating beyond the original model in logit space to surface rare behaviors.