method
active
method:per-paragraph-attribution-case-studyPer-paragraph attribution case study
Attribution share by system prompt paragraph relative to neutral baseline.
Neighborhood — ranked by edge-count
Papers (1)
paper
Frameworks (1)
framework
- Attribution graphs with nodes per turn per layer.
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.
- The task of attributing model behaviors to specific training datapoints.
- Measure causal importance of upstream variables.
- Correlating attribution vectors (feature activation × logit weight of next token) across model pairs to measure functional universality
- Gradient-based technique using SAE features to estimate causal effects on completions; used to corroborate NLA findings.
- Table 2; e.g., Identity +1.8 for turn-averaged.
- Gradient-based method to estimate the effect of zeroing a feature on a specific logit difference.
- Dot product of summed decoder vector with unembedding vector, following Hanna et al.
- Contrast in §6.3.