Robustness of Fine-Tuned Llms Under Noisy Retrieval Inputs
Yinghao Sang · 2025
Large Language Models (LLMs) have become foundational in natural language processing, particularly when finetuned for specific tasks. However, their effectiveness can diminish significantly when subjected to noisy or irrelevant retrieval inputs. This paper investigates the robustness of fine-tuned LLMs in retrieval-augmented generation systems, where noisy retrieval conditions are prevalent. We propose an end-to-end approach featuring user embedding and preference profiling, adaptive reward function design, and a soft prompt compression mechanism. We demonstrate significant robustness improvements via large-scale experiments on diverse datasets. A case study with detailed analysis also depicts the real-world benefits of our method. We conclude by outlining limitations, and ethics, and pointing out directions for future research.