Research
Sketching Algorithm Cuts Cost of Leave-One-Out Context Attribution in Language Models
This paper introduces a sketching algorithm that estimates leave-one-out (LOO) context attribution for language models, reducing the number of model evaluations from n to O(sqrt(log(1/epsilon))/epsilon) under a stability assumption. The authors prove a matching lower bound, showing the approach is near-optimal. Empirical tests on a small retrieval task confirm the method's effectiveness.
context attributionalgorithm efficiencymodel interpretability