PiSSA

Muhan Zhang's group (Institute for AI): SVD-initialized LoRA (NeurIPS 2024) — initializing adapters with the principal singular values and vectors of the frozen weight matrices converges faster and outperforms vanilla LoRA at identical cost. Merged upstream into Hugging Face PEFT (init_lora_weights="pissa"), making it one of the few academic PEFT variants with default-library distribution.