Authors: Edward J. Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen
Cite as: \cite{hu2021lora}
arXiv: https://arxiv.org/abs/2106.09685
Summary in three sentences
LoRA freezes pretrained model weights and injects trainable low-rank matrices into Transformer layers for downstream adaptation. The arXiv abstract presents LoRA as a parameter-efficient alternative to full fine-tuning with no additional inference latency after merging. For this project, LoRA is the main baseline search space that DoRA-style decomposed ES must beat.
Main contribution
The main contribution is Low-Rank Adaptation, a practical PEFT method for adapting large language models by training low-rank update matrices while keeping the base weights frozen.
Key equations and definitions
The key definition is a low-rank update added to a frozen pretrained weight matrix , usually with rank much smaller than the matrix dimensions. This gives the adapter coordinate system used by the plain LoRA ES baseline.
Empirical results
Every number here must name the section or figure it comes from.
If you cannot locate it in the paper, do not write it.
The arXiv abstract reports comparisons against full fine-tuning and adapter methods on several model families. Exact parameter, memory, throughput, and quality numbers should be copied only after reading the corresponding paper sections.
Relation to our project
- what we reuse: the frozen-weight low-rank adapter setup and plain LoRA baseline.
- what we dispute: LoRA coordinates may not be the best search space for reward-weighted ES perturbations.
- what we must reproduce: a minimal LoRA adapter implementation with matched rank, trainable parameter count, perturbation scale, and evaluation budget.
Criticism and limitations
LoRA was designed for gradient-based fine-tuning, not black-box reward optimisation. Its efficiency as a parameterisation does not guarantee that ES will search it effectively.