When Edit Localization Amplifies Relative Selection Bias: Gradient Geometry, Target Mismatch, and Importance Weighting

Title: When Edit Localization Amplifies Relative Selection Bias: Gradient Geometry, Target Mismatch, and Importance Weighting Source: arxiv Analyzing localized gradient updates reveals that edit localization can nonmonotonically alter relative selection bias depending on component biases and gradient geometry. The authors derive finite-sample mean-squared error and an optimal retention coefficient to mitigate target mismatch when using oracle importance weighting for localized objectives.