A new paper on arXiv examines how neural optimizers transfer across networks of different depths, proposing L-smoothness as a theoretical framework for the analysis. The authors argue that efficient scaling of deep learning models depends on predictable performance transfer across compute regimes.

The paper's central motivation is hyperparameter transfer: if training dynamics and structural choices discovered in one setting can be reliably carried over to another, researchers can avoid re-tuning at every scale. The title points to depth as the specific axis of transfer, with L-smoothness providing the mathematical condition that makes such transfer tractable.

The abstract is truncated, so the paper's specific results and experimental evidence are not available from the source. What is clear is the framing: depth transfer of optimizers is a problem worth solving, and L-smoothness is offered as a way to reason about it.