Identifying Specification Mismatch in Data Attribution Estimators
September 28, 2026
Influence estimators for data attribution often disagree due to specification mismatch rather than approximation error. The authors formalize influence as a counterfactual estimand, showing that choices in target behavior, intervention, and surrogate models fundamentally change the resulting rankings.
HOW THIS AFFECTS YOU
●
builderThis clarifies why different debugging tools may yield conflicting results when evaluating training data.
●
researcherYou must account for specification choices when comparing different data attribution methods.