Near-duplicate training samples lower a model's self-influence on a record while raising its extractability, so self-influence alone underestimates memorization risk.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Counterfactual Influence as a Distributional Quantity
Near-duplicate training samples lower a model's self-influence on a record while raising its extractability, so self-influence alone underestimates memorization risk.