RouteRAG

• 📚  https://arxiv.org/abs/2512.09487
• 👨🏽‍💻  https://github.com/YucanGuo/RouteRAG

This paper from UCAS presents RouteRAG, a reinforcement learning-based framework that enables LLMs to perform multi-turn RAG using both unstructured text and structured knowledge graphs. Unlike existing RAG systems that rely on fixed retrieval pipelines or single-shot retrieval, RouteRAG learns a unified policy that dynamically decides when to reason, what retrieval mode to use (passage, graph, or hybrid), and when to generate final answers. The system employs a two-stage training approach using Group Relative Policy Optimization: the first stage optimizes for answer correctness, while the second stage introduces an efficiency reward to minimize unnecessary retrieval overhead while maintaining accuracy. Experiments across five question-answering benchmarks demonstrate that RouteRAG significantly outperforms existing baselines, achieving comparable performance to GPT-4o-mini-based graph RAG systems despite using much smaller Qwen2.5 models (3B and 7B parameters), and surpassing the strongest RL-trained baseline (Search-R1) while using only 10k training examples compared to 170k.