The Hidden Power of Scaling Factor in LoRA Optimization

2026-06-11

AI Summary

This work challenges the common practice of treating the LoRA scaling factor α as merely an auxiliary to the learning rate, which has led to its underutilization. Through theoretical analysis and large-scale experiments, we demonstrate that α plays a dominant and distinct role—separate from the learning rate—in modulating task-relevant signals and mitigating gradient drift. Within a signal–drift framework, we show that α can effectively amplify task signals without exacerbating drift and derive a sublinear square-root relationship between α and the adapter rank. Building on these insights, we propose a LoRA-α minimization optimization framework that significantly enhances performance across diverse tasks, simplifies hyperparameter tuning, and enables LoRA to operate effectively with standard small learning rates while fully unlocking its learning capacity.

Abstract

In Low-Rank Adaptation (LoRA), the scaling factor $α$ is often treated as a mere complement to the learning rate, yet its role in optimization remains poorly understood. In this paper, we reveal that the scaling factor $α$ and the learning rate function differently, with $α$ emerging as the dominant driver of effective optimization, delivering gains that cannot be replicated by learning rate scaling alone. Through the synergy of extensive empirical analysis and a theoretical Signal-Drift framework, we uncover three findings into LoRA's scaling mechanism: First, LoRA's spectral suppression smooths the optimization landscape, rendering standard hyperparameters overly conservative and creating an optimization gap. Second, when leveraging this smoothness to accelerate convergence, $α$ outperforms the learning rate by amplifying the task signal without increasing the drift ratio. Third, the optimal scaling factor follows a sublinear relationship with the rank, well characterized by a square-root law with an unexpectedly large coefficient, revealing the insufficient scaling of existing rank-tied heuristics. Based on these insights, we propose LoRA-$α$, a minimalist framework that restores $α$ to its principled regime, making LoRA compatible with standard small learning rates. Extensive evaluations across diverse tasks demonstrate that LoRA-$α$ consistently improves performance while streamlining hyperparameter search, unleashing the learning potential of LoRA.

Problem

• LoRA
• scaling factor
• optimization
• hyperparameter
• low-rank adaptation

Innovation

• LoRA
• scaling factor
• optimization landscape
• signal-drift framework
• hyperparameter tuning