method
active
method:attribution-graphsAttribution Graphs
Gradient-based technique using SAE features to estimate causal effects on completions; used to corroborate NLA findings.
Neighborhood — ranked by edge-count
Papers (7)
paper
Frameworks (2)
framework
- MLP replacement with sparse features underlying the original attribution graphs.
- Attribution graphs with nodes per turn per layer.
Findings (2)
finding
- Mechanistic insight surfaced by NLA explanations and validated through independent causal attribution method.
- Illustrates NLA's capture of high-level cognition and hallucination of specifics; corroborated with attribution graphs.
Methods (6)
method
- Attribution graph constructionrelated_tosame_asMethod to trace how parameter subcomponents interact from input to output for a given next-token prediction, producing a subnetwork graph.
- Replacement model architecture with 30M features replacing MLP neurons.
- QK AttributionsextendsCore method decomposing attention scores as bilinear sums of query-key feature pairs.
- SupernodesusesManual grouping of related features to simplify attribution graphs.
- Dictionary used for all main-text features.
- Head LoadingsextendsPer-edge contribution of each attention head to attribution graph edges.
Artifacts (3)
artifact
- Attribution Graphs FrontendimplementsAnthropic open-source visualization tools used for graphs and max-activating examples.
- Open-source repo the authors hope to see extended with attention attributions.
- Interface to be extended with attentional attributions.
Claims (1)
claim
- Research philosophy.
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.
- Quantifies method limitations.
- Gradient-based method to estimate the effect of zeroing a feature on a specific logit difference.
- Correlating attribution vectors (feature activation × logit weight of next token) across model pairs to measure functional universality
- Shows how VPD-identified subnetworks can be analyzed to reveal interpretable pathways of computation (e.g., gender signal routing, syntactic role detection).
- Measure causal importance of upstream variables.
- Attribution graph reveals a pathway that detects the verb 'lost' and upweights object pronounsfinding0.794Second component of the subnetwork for 'her', complementing the femaleness signal.
- Baseline method against which probe-based ranking is compared; more computationally expensive.