aiqbench analysis
This study reveals that adaptive LoRA rank allocation, effective for supervised fine-tuning, actually degrades performance in Group Relative Policy Optimization (GRPO) by creating harmful gradient amplification loops. The authors demonstrate that reinforcement learning landscapes are fundamentally flatter than SFT settings, meaning all layers carry meaningful signal and uniform parameter budgets yield better accuracy under these conditions. This finding cautions against naively transferring efficiency strategies from pre-training or instruction tuning directly to alignment training without re-evaluating their impact on gradient dynamics.
When Gradient Importance Lies: Adaptive LoRA Rank Allocation Fails Under GRPO
arXiv:2605.07366v2 Announce Type: replace Abstract: Adaptive rank allocation for LoRA - allocating more parameters to important layers and fewer to unimportant ones - consistently improves efficiency under supervised fine-tuning (SFT). We test whether this success transfers to reinforcement learning, specifically Group Relative Policy Optimization (GRPO). Using gradient-magnitude profiling on Qwen 2.5 1.5B with GSM8K, we find that, in our setting, it does not: proportional rank allocation degrades accuracy by 4.5 points compared to uniform allocation (70.0% vs. 74.5%), despite using identical parameter budgets. We identify two mechanisms behind this failure. First, the gradient landscape under GRPO is fundamentally flatter than under SFT: the max-to-min layer importance ratio is only 2.17x, whereas the layer concentration reported by Shi et al. (2024) for SFT (top 30% of layers carrying >80% of the gradient signal) implies a max/min ratio well above 10x. All layers carry meaningful gradient signal; none are truly idle. Second, we observe a gradient amplification effect: non-uniform allocation widens the importance spread from 2.17x to 3.00x, creating a positive feedback loop where high-rank layers absorb more gradient while low-rank layers are progressively silenced. A random-allocation control yields the same amplification (r=0.972 correlation between assigned rank and resulting gradient share), indicating that rank causally determines gradient importance rather than the reverse. The negative result is single-seed and single-task; we present it as preliminary evidence that gradient importance does not predict capacity requirements under RL, and that naive transfer of SFT-era rank allocation strategies to alignment training should be evaluated cautiously.
Abstract:Adaptive rank allocation for LoRA - allocating more parameters to important layers and fewer to unimportant ones - consistently improves efficiency under supervised fine-tuning (SFT). We test whether this success transfers to reinforcement learning, specifically Group Relative Policy Optimization (GRPO). Using gradient-magnitude profiling on Qwen 2.5 1.5B with GSM8K, we find that, in our setting, it does not: proportional rank allocation degrades accuracy by 4.5 points compared to uniform allocation (70.0% vs. 74.5%), despite using identical parameter budgets. We identify two mechanisms behind this failure. First, the gradient landscape under GRPO is fundamentally flatter than under SFT: the max-to-min layer importance ratio is only 2.17x, whereas the layer concentration reported by Shi et al. (2024) for SFT (top 30% of layers carrying >80% of the gradient signal) implies a max/min ratio well above 10x. All layers carry meaningful gradient signal; none are truly idle. Second, we observe a gradient amplification effect: non-uniform allocation widens the importance spread from 2.17x to 3.00x, creating a positive feedback loop where high-rank layers absorb more gradient while low-rank layers are progressively silenced. A random-allocation control yields the same amplification (r=0.972 correlation between assigned rank and resulting gradient share), indicating that rank causally determines gradient importance rather than the reverse. The negative result is single-seed and single-task; we present it as preliminary evidence that gradient importance does not predict capacity requirements under RL, and that naive transfer of SFT-era rank allocation strategies to alignment training should be evaluated cautiously.
Computation and Language (cs.CL)
arXiv:2605.07366 [cs.CL]
(or arXiv:2605.07366v2 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2605.07366 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 337 words
Related coverage
Statistical Inference for Rank Allocation in Low-Rank Adaptation
arXiv · Jul 23, 2026
IFCLoRA: Topology-Aware Rank Allocation for Parameter-Efficient Fine-Tuning
arXiv · Jul 28, 2026
LoRA-GA$^2$: Low Rank Adaptation with Multi-step Gradient Adaptive Alignment
arXiv · Aug 21, 2026
Train Large, Deploy Compact: Structured Compression for Compact Low-Rank Adaptation
arXiv · Jul 31, 2026