SST vs LoRA: A Leaner, Smarter Way to Train AI Models

This study introduces SST, a low-rank optimization method that achieves near full-rank performance in AI model training while drastically reducing trainable parameters. Tested on the OPT language model and Hyperbolic Graph Neural Networks (HGNNs), SST outperformed LoRA and ReLoRA across multiple benchmarks — from zero-shot NLP evaluations to node classification and link prediction. The results show that SST offers a more efficient and scalable alternative for training large models without sacrificing accuracy or generalization.