Research map / Architectures and efficiency / Pruning and efficient tuning
Rethinking LoRA Initialization for Robust Asymmetric Learning Rates
Disen Liao, Yaoliang Yu
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Revisits LoRA initialization to make asymmetric learning-rate training (as in LoRA+) more robust, reducing the cost and brittleness of LR selection.
Problem
LoRA is highly sensitive to learning-rate choices, and tuning LRs is costly; LoRA+ uses asymmetric LRs for the two low-rank factors but the abstract indicates the practical question is reducing LR-selection brittleness.
Approach
Rethinks LoRA initialization as the lever for making asymmetric learning rates robust, rather than designing new LoRA variants. Full method details are not available in the provided abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| initialize A with variance n^-2 (Init[A^2]) instead of n^-1 (Init[A]) | LoRA with asymmetric LRs · MNIST toy MLP, beta=5.0, width 4096 | validation loss and optimal LR transfer | Improved |
| initialize A with variance n^-2 instead of n^-1 | LoRA, Llama 3-8B · arithmetic and commonsense, varying beta | accuracy across LR ratios | Improved |
| initialize A with variance n^-2 instead of n^-1 | LoRA, Llama 3.2-1B · arithmetic and commonsense, varying rank r | accuracy across LR ratios | Improved |
| raise LoRA multiplier alpha/r from 3 to 12 under Init[A] (ablation) | Init[A], Llama 3-8B · arithmetic | accuracy (%): 76.4 -> 72.0 | Hurt |
| raise LoRA multiplier up to alpha/r=10-12 under Init[A^2] (ablation) | Init[A^2], Llama 3-8B · arithmetic | accuracy (%): 77.5 at alpha/r=10; 77.3 at alpha/r=12 | Improved |
| use Init[A^2] instead of Init[AB] with equal learning rates | LoRA · LLM fine-tuning | performance | Improved |
| vary rank r under Init[A^2] (ablation) | Init[A^2], Llama 3.2-1B · math training loss | optimal LR | Mixed |
| suppress A variance further (n^-3, n^-4) (ablation) | Init[A^2] · theory | noise suppression | Improved |
Experimental setup
Baselines: vanilla LoRA LoRA+
Builds on: LoRA LoRA+
Closest papers by meaning
- The Impact of Initialization on LoRA Finetuning Dynamics (NeurIPS 2024 · similarity 0.93)
- LoRA Training Provably Converges to a Low-Rank Global Minimum Or It Fails Loudly (But it Probably Won't Fail) (ICML 2025 · similarity 0.92)
- Beyond Zero Initialization: Investigating the Impact of Non-Zero Initialization on LoRA Fine-Tuning Dynamics (ICML 2025 · similarity 0.92)
- LoRA-GA: Low-Rank Adaptation with Gradient Approximation (NeurIPS 2024 · similarity 0.91)
- Learning Rate Matters: Vanilla LoRA May Suffice for LLM Fine-tuning (NeurIPS 2026 · similarity 0.90)
- The Primacy of Magnitude in Low-Rank Adaptation (NeurIPS 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
