DACL-RAG: Data Augmentation Strategy with Curriculum Learning for Retrieval-Augmented Generation

TL;DR

DACL-RAG tackles noise and limited discriminability in retrieval-augmented generation by introducing a multi-level data augmentation strategy and a multi-stage curriculum that jointly optimizes the generator and retriever. The data evolution framework creates easy-to-hard training sets via document rewriting and controlled noise, while curriculum stages progressively align the components with a well-trained RALM. Empirical results on four open-domain QA datasets show consistent 2-4% gains over strong baselines and improved robustness and generalization across model scales and out-of-domain data. This work demonstrates that integrating curriculum learning with data augmentation can stabilize and elevate RAG systems, enabling more reliable retrieval-guided language understanding and generation.

Abstract

Retrieval-Augmented Generation (RAG) is an effective method to enhance the capabilities of large language models (LLMs). Existing methods typically optimize the retriever or the generator in a RAG system by directly using the top-k retrieved documents. However, two key issues inherent in the training data constrain the effectiveness of this training paradigm: (1) across different queries, the top-k retrieved documents vary greatly in content quality, with some providing valuable knowledge while others lack critical information or are even misleading, and training on such data in a purely random manner may impair the generator's ability to extract key information; (2) for a given query, the limited set of k documents often exhibits low discriminability, and training solely on them makes it difficult for the retriever to learn how to distinguish between relevant and irrelevant documents. To address these issues, we introduce DACL-RAG, a multi-stage RAG training framework that combines a multi-level Data Augmentation strategy with a multi-stage Curriculum Learning paradigm. The data augmentation strategy constructs comprehensive and diverse training sets with controllable difficulty levels through sample evolution, while the curriculum learning paradigm organizes them into progressive stages for training, ensuring stable and consistent improvements, thereby optimizing the overall performance and generalization of the RAG system more effectively. Our DACL-RAG framework demonstrates consistent effectiveness across four open-domain QA datasets, achieving performance gains of 2% to 4% over multiple advanced methods.

Figures (8)

• Figure 1: Existing training paradigm. Here, green denotes documents that support the model’s responses, while red denotes documents that are useless or even harmful. We conduct analysis on four open-domain QA datasets: NQ, TriviaQA, HotpotQA, and PopQA. For generator training (upper panel): Raw Retrieve Result reports the answer recall of the original retrieved documents; these raw retrievals exhibit low recall of answer-bearing evidence, with some documents containing the answer and others not, so training directly on such inputs injects noise and impedes optimization. For retriever training (lower panel): after scoring document usefulness with an LLM, we find that within the $top$-$5$ candidates the ranking scores exhibit limited dispersion, providing insufficient signal to learn robust discrimination between helpful and unhelpful documents; expanding the sampling window to the $top$-$20$ increases score separation and mitigates this issue.
• Figure 2: The overview of our DACL-RAG training framework, which has into two continuous phases: (1) Training Generator: We construct multiple difficulty levels of documents and then finetune a RALM in a stage-by-stage manner. (2) Training Retriever: We use the well-trained RALM to assess documents and rerank them. We then construct document data from easy to difficult, and finetune the retriever in a stage-by-stage manner.
• Figure 3: Case of document rewriting, including query-enhanced document rewriting and counterfactual document rewriting. The highlighted green text represents the keywords within the document that aid in addressing the question, while the highlighted red text signifies wrong knowledge.
• Figure 4: Generalization experiments of DACL-RAG training methods across Qwen2.5-3B model, we report the average Exact Match (EM) scores for generation accuracy evaluation and Recall@k scores for retrieval quality assessment, which demonstrate consistent superiority of our approach.
• Figure 5: Evaluation results on the out-of-domain dataset 2WikiMultiHopQA, where EM and F1 assess answer accuracy, and R@k evaluates the quality of retrieved documents.