VeRA

VeRA: Vector-based Random Matrix Adaptation (ICLR 2024) — a parameter-efficient fine-tuning method that drastically cuts trainable parameters versus LoRA by sharing frozen random matrices across layers and training only small scaling vectors; adopted into HuggingFace PEFT. QUVA Lab, UvA. The arXiv preprint predates the tracking window, but VeRA is retained as a foundational, widely-adopted PEFT technique.