This paper introduces Retrieval-augmented Adaptive Adversarial Training (RAAT), a novel approach to enhance the noise robustness of Retrieval-Augmented Language Models (RALMs). RALMs face challenges due to retrieval noises, which can lead to incorrect responses. The authors categorize retrieval noises into three types: relevant retrieval noise (superficially related but lacking correct answers), irrelevant retrieval noise (low relevance), and counterfactual retrieval noise (topically related but containing incorrect information). They propose RAAT, which uses adaptive adversarial training to dynamically adjust the model's training process in response to retrieval noises and integrates multi-task learning to enable the model to internally recognize noisy contexts. RAAT generates adversarial samples based on the model's sensitivity to different noise types, aligning with the min-max paradigm of adversarial training. The method also incorporates a regularization term to reduce the model's sensitivity to retrieval noise by encouraging balanced optimization. Experiments on three open-domain question-answering datasets show that RAAT significantly improves the F1 and EM scores of the LLaMA-2 7B model under various noise conditions. The authors also establish a benchmark (RAG-Bench) to evaluate the noise robustness of RALMs. Results demonstrate that RAAT outperforms existing methods in handling diverse retrieval noise environments. The study highlights the importance of addressing retrieval noise in RALMs to improve their robustness and reliability.This paper introduces Retrieval-augmented Adaptive Adversarial Training (RAAT), a novel approach to enhance the noise robustness of Retrieval-Augmented Language Models (RALMs). RALMs face challenges due to retrieval noises, which can lead to incorrect responses. The authors categorize retrieval noises into three types: relevant retrieval noise (superficially related but lacking correct answers), irrelevant retrieval noise (low relevance), and counterfactual retrieval noise (topically related but containing incorrect information). They propose RAAT, which uses adaptive adversarial training to dynamically adjust the model's training process in response to retrieval noises and integrates multi-task learning to enable the model to internally recognize noisy contexts. RAAT generates adversarial samples based on the model's sensitivity to different noise types, aligning with the min-max paradigm of adversarial training. The method also incorporates a regularization term to reduce the model's sensitivity to retrieval noise by encouraging balanced optimization. Experiments on three open-domain question-answering datasets show that RAAT significantly improves the F1 and EM scores of the LLaMA-2 7B model under various noise conditions. The authors also establish a benchmark (RAG-Bench) to evaluate the noise robustness of RALMs. Results demonstrate that RAAT outperforms existing methods in handling diverse retrieval noise environments. The study highlights the importance of addressing retrieval noise in RALMs to improve their robustness and reliability.