AI News

When Gradient Masks Lose: Random Density-Matched Masks Beat Gradient Attribution in LoRA B-Matrix Forgetting Tests

ABC2 min read
Research