Zeroth-order SGD learning dynamics are governed by a random low-dimensional projection of the empirical NTK whose approximation error scales with model output dimension, not parameter count.
arXiv preprint arXiv:2505.18886 , year=
2 Pith papers cite this work. Polarity classification is still indexing.
2
Pith papers citing it
fields
cs.LG 2representative citing papers
Downgrading optimizers to lower-information variants during LLM unlearning yields more robust forgetting on MUSE and WMDP benchmarks by converging to harder-to-perturb loss basins.
citing papers explorer
-
Learning Dynamics of Zeroth-Order Optimization: A Kernel Perspective
Zeroth-order SGD learning dynamics are governed by a random low-dimensional projection of the empirical NTK whose approximation error scales with model output dimension, not parameter count.
-
Downgrade to Upgrade: Optimizer Simplification Enhances Robustness in LLM Unlearning
Downgrading optimizers to lower-information variants during LLM unlearning yields more robust forgetting on MUSE and WMDP benchmarks by converging to harder-to-perturb loss basins.