Enhancing Noise Robustness of Retrieval-Augmented Language Models with Adaptive Adversarial Training (RAAT)
- Paper: https://arxiv.org/abs/2405.20978
- Authors: Feiteng Fang, Yuelin Bai, Shiwen Ni, Min Yang, Xiaojun Chen, Ruifeng Xu
- Published: 2024-05-31
- Code: https://github.com/calubkk/RAAT
- Topic area: RAG robustness under retrieval noise (LLMs and large-scale search)
- Problem: RAG mitigates LLM hallucination, outdated knowledge, and untraceable reasoning, but inappropriate retrieved passages can degrade response quality. Prior noise-robustness studies consider only a limited set of noise types, deviating from real-world retrieval environments.
- Noise taxonomy: The authors investigate retrieval noises and categorize them into three distinct types that reflect real-world retrieval conditions, then analyze how each type affects LLM robustness.
- Method — RAAT (Retrieval-augmented Adaptive Adversarial Training):
- Uses adaptive adversarial training to dynamically adjust the model's training process in response to retrieval noises.
- Employs multi-task learning to ensure the model can internally recognize noisy contexts.
- Results: A LLaMA-2 7B model trained with RAAT shows significant improvements in F1 and EM scores under diverse noise conditions.
- Reproducibility: Code and data are publicly released.
- Quantitative results should be verified against the original PDF tables; this page is based on the abstract and public metadata.
- Experiments center on LLaMA-2 7B; cross-lingual and larger-model generalization remain open questions.
- Latency, cost, and safety constraints of industrial deployment are not the paper's focus.
- FreshLLMs: Refreshing Large Language Models with Search Engine Augmentation (arXiv:2310.03214)
- Interleaving Retrieval with Chain-of-Thought Reasoning (arXiv:2212.10509)
- Long-form factuality in large language models (arXiv:2403.18802)
- ReSearch: Learning to Reason with Search via Reinforcement Learning (arXiv:2503.19470)
- Search-R1: Training LLMs to Reason and Leverage Search Engines (arXiv:2503.09516)
Key points
Context and significance
The work sits at the intersection of RAG and large-scale search systems, addressing how to redistribute responsibilities among retrieval, ranking, and generation in the LLM era. Its findings are relevant to:
1. Architecture: cascaded retrieve-rerank-generate pipelines remain mainstream, but noise-aware training makes robustness an learnable property. 2. Evaluation: robustness should be measured under diverse, realistic noise conditions rather than a single perturbation type. 3. Deployment: recognizing noisy contexts internally helps production systems avoid over-trusting retrieval output.
Limitations noted by the community review
Original abstract
> Large Language Models (LLMs) exhibit substantial capabilities yet encounter challenges, including hallucination, outdated knowledge, and untraceable reasoning processes. Retrieval-augmented generation (RAG) has emerged as a promising solution, integrating knowledge from external databases to mitigate these challenges. However, inappropriate retrieved passages can potentially hinder the LLMs' capacity to generate comprehensive and high-quality responses. Prior RAG studies on the robustness of retrieval noises often confine themselves to a limited set of noise types, deviating from real-world retrieval environments and limiting practical applicability. In this study, we initially investigate retrieval noises and categorize them into three distinct types, reflecting real-world environments. We analyze the impact of these various retrieval noises on the robustness of LLMs. Subsequently, we propose a novel RAG approach known as Retrieval-augmented Adaptive Adversarial Training (RAAT). RAAT leverages adaptive adversarial training to dynamically adjust the model's training process in response to retrieval noises. Concurrently, it employs multi-task learning to ensure the model's capacity to internally recognize noisy contexts. Extensive experiments demonstrate that the LLaMA-2 7B model trained using RAAT exhibits significant improvements in F1 and EM scores under diverse noise conditions. For reproducibility, we release our code and data at: https://github.com/calubkk/RAAT.