From What Is Used to What Is Missing: Rethinking Dynamic Rank in LoRA
Abstract
Parameter-efficient fine-tuning with LoRA is governed by rank, which controls parameter budget and locally accessible task updates. Existing dynamic-rank methods mainly focus on capacity allocation, leaving unresolved whether the current adaptation space is sufficient for the task. We characterize the first-order information boundary of standard factor gradients, showing that their unobservable subspace coincides with the normal space of the low-rank manifold. Thus, factor gradients alone cannot distinguish under-optimization within the current space from task-relevant directions beyond its first-order reach. This motivates a geometric view of rank evolution, where tangent information optimizes existing degrees of freedom and the normal spectrum quantifies the value of new ones. Under a local quadratic model, we show that leading normal singular directions yield the locally optimal rank expansion, with spectral energy determining its gain. Building on this characterization, we propose GEAR, which turns the tangent--normal decomposition into rank evolution through tangent-space optimization, matrix-free normal-spectrum estimation for rank expansion, and task-induced geometric initialization. Across language and vision benchmarks, GEAR achieves the best average performance in all evaluated settings with substantially fewer trainable parameters. On LLaMA3-8B, it raises the average commonsense score from 86.61 to 88.45 while reducing trainable parameters from 42.0M to 14.4M.
Related
- What does ethnic density represent? Spatial co-occurrence networks of a widely used contextual measure using harmonised UK small-area census data
- Three Quadratic Recurrences Modulo 3: What We Can Prove and What We Can Guess
- Transition to an oil-based economy in Senegal: What are the expected effects on external macroeconomic dynamics?
- What masking geometry works best for EEG foundation models?
- What professional knowledge of mathematics do CLIS teachers need?
BibTeX
@misc{xue2026from,
title = {{From What Is Used to What Is Missing: Rethinking Dynamic Rank in LoRA}},
author = {Peng Xue and Guodong Fan and Guang-Yong Chen and Xing Chen},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.07349},
eprint = {2611.07349},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.07349},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - From What Is Used to What Is Missing: Rethinking Dynamic Rank in LoRA
AU - Peng Xue
AU - Guodong Fan
AU - Guang-Yong Chen
AU - Xing Chen
PY - 2026
DA - 2026/11/05
PB - arXiv
DO - 10.48550/arxiv.2611.07349
UR - https://arxiv.org/abs/2611.07349
AB - Parameter-efficient fine-tuning with LoRA is governed by rank, which controls parameter budget and locally accessible task updates. Existing dynamic-rank methods mainly focus on capacity allocation, leaving unresolved whether the current adaptation space is sufficient for the task. We characterize the first-order information boundary of standard factor gradients, showing that their unobservable subspace coincides with the normal space of the low-rank manifold. Thus, factor gradients alone cannot distinguish under-optimization within the current space from task-relevant directions beyond its first-order reach. This motivates a geometric view of rank evolution, where tangent information optimizes existing degrees of freedom and the normal spectrum quantifies the value of new ones. Under a local quadratic model, we show that leading normal singular directions yield the locally optimal rank expansion, with spectral energy determining its gain. Building on this characterization, we propose GEAR, which turns the tangent--normal decomposition into rank evolution through tangent-space optimization, matrix-free normal-spectrum estimation for rank expansion, and task-induced geometric initialization. Across language and vision benchmarks, GEAR achieves the best average performance in all evaluated settings with substantially fewer trainable parameters. On LLaMA3-8B, it raises the average commonsense score from 86.61 to 88.45 while reducing trainable parameters from 42.0M to 14.4M.
ER -