Data Attribution

The task of attributing model behaviors to specific training datapoints.

Neighborhood — ranked by edge-count

paper

community

method

Probe-Based Data Attribution
implements
Linear classifier approach applied to model activations to identify which training datapoints caused undesired behaviors in post-training.

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Attribution Similaritymethod0.850
Correlating attribution vectors (feature activation × logit weight of next token) across model pairs to measure functional universality
Gradient-based data attributionmethod0.834
Baseline method against which probe-based ranking is compared; more computationally expensive.
Attribution Graphsmethod0.834
Gradient-based technique using SAE features to estimate causal effects on completions; used to corroborate NLA findings.
Attribution patchingmethod0.802
Gradient-based method to estimate the effect of zeroing a feature on a specific logit difference.
Probe-based data attribution for alignmentconcept0.798
Attribution graph constructionmethod0.780
Method to trace how parameter subcomponents interact from input to output for a given next-token prediction, producing a subnetwork graph.
Attribution of agency is an empirical question, not a philosophical one.claim0.767
LLM-Judge Data Attributionmethod0.765
Alternative data attribution approach using an LLM as a judge; compared against the probe-based method.