LoRA-Muon: New Optimizer Improves Low-Rank Adaptation for Fine-Tuning Deep Learning Models
Researchers have developed LoRA-Muon, a new optimizer that applies spectral steepest-descent principles to low-rank adaptation (LoRA), a technique for efficiently fine-tuning large language models. The method addresses known challenges with LoRA, including sensitivity to initialization and poor transfer of learning rates across different ranks. The work is significant because it demonstrates that low-rank fine-tuning can match or exceed the performance of full-rank training while maintaining computational efficiency.
LoRA-Muon is a new optimization algorithm designed to improve low-rank adaptation, a popular technique for reducing computational and memory costs when fine-tuning deep learning models. The researchers derived LoRA-Muon by applying the Muon optimizer's spectral steepest-descent rule to the low-rank setting, combined with a novel split weight-decay approach. Their key contribution is demonstrating that LoRA-Muon serves as an effective low-rank proxy for full-rank optimizers like Muon and Shampoo-family methods. In experiments on TinyShakespeare, the method shows that optimal learning rates transfer reliably across different ranks and model dimensions, and that rank-32 LoRA-Muon achieves lower validation loss than dense baselines in seed-averaged sweeps. Additionally, the algorithm avoids expensive QR-decomposition and second-moment storage, making it more memory-efficient and accelerator-friendly than competing approaches.
What's missing
The paper does not provide extensive empirical validation on large-scale models or diverse downstream tasks beyond TinyShakespeare; broader evaluation on modern language models and diverse fine-tuning scenarios would strengthen claims about generalizability. The authors do not discuss computational wall-clock time comparisons or provide code availability statements.
What different sources said
- arXiv cs.AICenter
LoRA-Muon: Spectral Steepest Descent on the Low-Rank Manifold
Related
Topology-Aware Thermodynamics Improves DNA Probe Specificity Design
Researchers developed a new framework for designing DNA probes that accounts for the spatial organization of matched sequences, not just overall thermodynamic stability. Traditional methods rely on scalar measures like melting temperature and free energy, which miss how mismatches are distributed along the probe. The approach could improve diagnostic accuracy in applications like HPV detection and gene expression profiling.
Study Identifies Optimal Thermal Dose for Combining Focused Ultrasound with Immunotherapy in Tumors
Researchers used multimodal PET imaging to identify an optimal thermal dose range for focused ultrasound ablation that destroys tumor tissue while preserving conditions for immunotherapy delivery. The study found that excessive heating collapses blood vessels needed for antibody access, while insufficient heating fails to adequately reduce tumor burden. The findings could guide clinical design of combination treatments pairing thermal ablation with immunotherapies.
Plant MSH1 Protein Functions as Mismatch-Directed Nuclease for Organelle Genome Maintenance
Researchers have identified the precise mechanism by which the AtMSH1 protein in Arabidopsis plants recognizes and cleaves DNA mismatches and lesions, preventing mutations in organellar genomes. The protein combines a DNA mismatch recognition module with a nuclease domain that makes staggered cuts at specific positions relative to DNA damage. This discovery explains how plants maintain unusually low mutation rates in their mitochondrial and chloroplast DNA compared to other eukaryotes.