Reproduce “VaccineRAG: Boosting Multimodal Large Language Models' Immunity to Harmful RAG Samples” on OpenResearch

About this paper

Qixin Sun, Ziqin Wang, Hengyuan Zhao, Yilin Li, Kaiyou Song, Linjiang Huang, Xiaolin Hu, Qingpei Guo and 1 more

VaccineRAG, from Beihang University and QIYUAN LAB, enhances multimodal large language models' resilience to harmful retrieved samples in Retrieval-Augmented Generation (RAG) systems. This is achieved through a Chain-of-Thought (CoT) based dataset and a new Partial-GRPO optimization, reducing accuracy degradation rates (e.g., from 42.45% to 10.43% for Qwen2.5-VL) while maintaining overall accuracy.

Key insights

• Explicit, fine-grained Chain-of-Thought reasoning can significantly improve MLLMs' ability to critically evaluate and utilize retrieved evidence, making them inherently more robust to noisy inputs.
• Targeted, segment-specific reward functions and localized gradient backpropagation (Partial-GRPO) are essential for efficiently training LLMs on complex, multi-component CoT outputs, overcoming limitations of uniform reward application.
• Strengthening the MLLM generator's internal immunity to harmful samples is a crucial complementary approach to solely improving retriever precision, providing a more reliable solution for real-world RAG systems.

Summary by alphaXiv.