New LoRA method better handles nonlinear gate activations in transformers

Original: Nonlinearity-Aware LoRA: Structured Gate Adaptation under Low-Rank Constraints

Source: arxiv.org ↗