RAAT: Adaptive Adversarial Training for Noise-Robust Retrieval-Augmented Language Models
Overview
| Field | Detail | |---|---| | Paper | Enhancing Noise Robustness of Retrieval-Augmented Language Models with Adaptive Adversarial Training | | Authors | Feiteng Fang, Yuelin Bai, Shiwen Ni, Min Yang, Xiaojun Chen, Ruifeng Xu | | Published | 2024-05-31 | | Source | https://arxiv.org/abs/2405.20978 | | Code | https://github.com/calubkk/RAAT |
Motivation
Large Language Models (LLMs) exhibit substantial capabilities yet encounter challenges including hallucination, outdated knowledge, and untraceable reasoning processes. Retrieval-augmented generation (RAG) mitigates these issues by integrating external knowledge, but inappropriate retrieved passages can hinder the LLM's ability to produce comprehensive, high-quality responses. Prior robustness studies often consider only a limited set of noise types, deviating from real-world retrieval environments.
Contributions
- Noise taxonomy: Categorizes retrieval noise into three distinct types that reflect real-world retrieval environments, and analyzes how each affects LLM robustness.
- RAAT method: Proposes Retrieval-augmented Adaptive Adversarial Training, which dynamically adjusts the model's training process in response to retrieval noises via adaptive adversarial training.
- Noise recognition: Employs multi-task learning so the model can internally recognize noisy contexts.
- Empirical results: A LLaMA-2 7B model trained with RAAT shows significant F1 and EM improvements under diverse noise conditions.
- Original paper: https://arxiv.org/abs/2405.20978
- Code and data: https://github.com/calubkk/RAAT
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 Large Language Models (LLMs) exhibit substantial capabilities yet encounter challenges, including hallucination, outdated knowledge, and untraceable reasoning processes.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.
Context in the RAG Landscape
RAAT sits at the intersection of RAG and large-scale search systems, addressing how to reallocate responsibility among retrieval, ranking, generation, and tool use in the LLM era. Evaluation is shifting from static ranking metrics (nDCG, MRR) toward task success rate, citation accuracy, and multi-hop reasoning chain completeness. Related entries worth reading alongside this one include FreshLLMs, Gorilla, IRCoT, ReSearch, and Search-R1.
Takeaways
1. Architecture: Cascade retrieval + reranking + generation remains mainstream, but agentic paradigms treat retrieval strategy itself as learnable. 2. Robustness matters: Real-world retrieval is noisy; training-time defenses like RAAT improve resilience without relying on perfect retrieval. 3. Evaluation: Offline metric gains should be cross-validated with human assessment before concluding production readiness.