V-RAG: Competitive Tree Reranking and Static Distillation for Answer-Source Alignment

Retrieval-Augmented Generation (RAG) has become a mainstream paradigm for enhancing the factual consistency of large language models (LLMs). However, existing RAG frameworks still face two major challenges: improving retrieval recall to provide more relevant context for generation, and reducing hallucinations in LLM-generated answers. In this paper, we propose V-RAG, a novel RAG framework augmented by a validation mechanism. In the reranking stage, V-RAG employs a multi-round filtering process based on a competition tree structure. An unfinetuned LLM is used to eliminate redundant documents from the initially retrieved set, enhancing contextual relevance while reducing computational overhead. In the generation stage, V-RAG introduces a three-way answer evaluation module that classifies responses into acceptable, hallucinated, or irrelevantly cited. It also dynamically updates prompt strategies based on feedback. To improve deployment efficiency, we further distill the 0.6 B -scale validator into a lightweight BERT-based classifier. We conducted experiments on two Chinese datasets: the CRUD_data and the Chinese tax and finance QA dataset. The results demonstrate that V-RAG outperforms multiple baselines in terms of recall and answer similarity (generated answer vs. standard answer), without requiring large-scale model fine-tuning.