A new arXiv paper, First-Order LoRA Training with Phone NPUs, targets on-device fine-tuning of large language models. The motivation is privacy: the model can learn from a user's data without uploading it. LoRA is used to reduce trainable parameters, making fine-tuning more feasible on a phone's neural processing unit.
The abstract cautions that LoRA does not eliminate all costs. While it reduces trainable parameters, each first-order training step still multiplies by something; the sentence is cut off in the available abstract. The paper's exact claim about that remaining cost cannot be determined from the announcement alone.
Because this digest is based on a single arXiv abstract, there are no independent sources to compare. The main takeaway is the paper's framing: LoRA helps with memory and parameter count, but first-order training on phone NPUs still has a bottleneck that the full text presumably addresses.