Improving text processing via adversarial low-rank adaptation

Hao Wu, Xiangfeng Luo, Jianqi Gao, Dian Huang

Abstract: Parameter-efficient fine-tuning (PEFT) has emerged as a predominant approach for adapting large language models(LLMs) to downstream natural language processing tasks. This method achieves significant savings in time and computational resources by training a small set of adapters parameter while preserving the main structure of the LLMs. Among various PEFT methods, low-rank adaptation (LoRA) is one of the widely used methods for fine-tuning LLMs. However, the parameter scale of the adapters is significantly smaller than that of the LLMs. This discrepancy may introduce notable instability when the model adapts to various downstream tasks, thereby potentially limiting the performance of LoRA. To address this issue, we propose an adversarial training-based low-rank adaptation method, termed ADV-LoRA. The core idea of this method is to increase the model’s instability by introducing adversarial perturbations and then reduce this instability through an adversarial training mechanism. Specifically, ADV-LoRA applies adversarial perturbations at the adapter parameter layer, causing a temporary decline in the performance of the LLMs on specific tasks. Subsequently, the adversarial training process gradually mitigates performance fluctuations, encouraging the LLMs to exhibit more consistent and stable performance on specific task data, thereby enhancing the robustness of model fine-tuning. Extensive experiments have demonstrated the effectiveness of ADV-LoRA, particularly in text processing tasks such as text classification and text generation.

External IDs:dblp:journals/ml/WuLGH25