Reproduce “EviNote-RAG: Enhancing RAG Models via Answer-Supportive Evidence Notes” on OpenResearch

About this paper

Yuqin Dai, Guoqing Wang, Yuan Wang, Kairan Dou, Kaichen Zhou, Zhanwei Zhang, Shuo Yang, Fei Tang and 8 more

EviNote-RAG is a reinforcement learning framework that enhances Retrieval-Augmented Generation (RAG) models through a "retrieve–note–answer" workflow, incorporating answer-supportive evidence notes (SENs) and an entailment-based Evidence Quality Reward (EQR). The approach achieves substantial accuracy gains, including a relative F1 gain of 20% on HotpotQA and over 91% F1 on 2Wiki, while also demonstrating improved training stability and efficiency.

Key insights

• Structuring the RAG pipeline with an explicit note-taking step (SENs) improves an LLM's ability to filter noise and concentrate on relevant evidence, enhancing multi-hop reasoning.
• Semantic supervision on intermediate reasoning steps via an entailment-based reward (EQR) is crucial for guiding the model to generate logically consistent evidence notes, leading to more stable and effective policy learning.
• The combination of structured evidence notes and fine-grained semantic rewards substantially boosts model generalization and robustness, particularly on out-of-domain and noisy datasets.

Summary by alphaXiv.