How Much is Too Much? Exploring LoRA Rank Trade-offs for Retaining Knowledge and Domain Robustness

TL;DR

This paper evaluates how different LoRA ranks affect the performance, generalization, and internal representations of large language models in downstream tasks, comparing them to full fine-tuning methods.

Contribution

It provides a comprehensive analysis of LoRA rank trade-offs, revealing optimal configurations for reasoning and recall tasks and insights into internal model representations.

Findings

LoRA can outperform full fine-tuning at certain ranks on reasoning tasks.

Optimal LoRA ranks balance knowledge retention and domain robustness.

Spectral and attention analyses reveal internal representational changes.

Abstract

Large language models are increasingly adapted to downstream tasks through fine-tuning. Full supervised fine-tuning (SFT) and parameter-efficient fine-tuning (PEFT) methods, such as Low-Rank Adaptation (LoRA), are two dominant approaches. While PEFT methods are widely used for their computational efficiency, the implications of their configurations (e.g., rank) remain under-explored in downstream Q&A tasks and generalisation. In this work, we perform a comprehensive evaluation across multiple reasoning and recall datasets, conducting a rank sweep to quantify the trade-off between SFT and PEFT. We also compare the accuracy of PEFT and SFT models across in-domain and out-of-domain adaptation, highlighting distinct generalisation behaviour and task-specific forgetting. We demonstrate that LoRA achieves competitive and in some cases superior performance compared to SFT, particularly on reasoning tasks at specific rank values. Additionally, we analyze the internal representations via spectral features and layer-wise attention structures, offering insights into representational drift and structural changes in attention patterns.